Tf model outputs (#6247)

* TF outputs and test on BERT

* Albert to DistilBert

* All remaining TF models except T5

* Documentation

* One file forgotten

* TF outputs and test on BERT

* Albert to DistilBert

* All remaining TF models except T5

* Documentation

* One file forgotten

* Add new models and fix issues

* Quality improvements

* Add T5

* A bit of cleanup

* Fix for slow tests

* Style

docs/source/model_doc/albert.rst

@@ -50,7 +50,10 @@ AlbertTokenizer
 Albert specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: transformers.modeling_albert.AlbertForPretrainingOutput
+.. autoclass:: transformers.modeling_albert.AlbertForPreTrainingOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_albert.TFAlbertForPreTrainingOutput
     :members:
 
 

docs/source/model_doc/bert.rst

@@ -57,7 +57,10 @@ BertTokenizerFast
 Bert specific outputs
 ~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: transformers.modeling_bert.BertForPretrainingOutput
+.. autoclass:: transformers.modeling_bert.BertForPreTrainingOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_bert.TFBertForPreTrainingOutput
     :members:
 
 

docs/source/model_doc/electra.rst

@@ -74,7 +74,10 @@ ElectraTokenizerFast
 Electra specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: transformers.modeling_electra.ElectraForPretrainingOutput
+.. autoclass:: transformers.modeling_electra.ElectraForPreTrainingOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_electra.TFElectraForPreTrainingOutput
     :members:
 
 
@@ -106,6 +109,13 @@ ElectraForSequenceClassification
     :members:
 
 
+ElectraForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ElectraForMultipleChoice
+    :members:
+
+
 ElectraForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -141,6 +151,20 @@ TFElectraForMaskedLM
     :members:
 
 
+TFElectraForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFElectraForSequenceClassification
+    :members:
+
+
+TFElectraForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFElectraForMultipleChoice
+    :members:
+
+
 TFElectraForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

docs/source/model_doc/gpt.rst

@@ -77,6 +77,9 @@ OpenAI specific outputs
 .. autoclass:: transformers.modeling_openai.OpenAIGPTDoubleHeadsModelOutput
     :members:
 
+.. autoclass:: transformers.modeling_tf_openai.TFOpenAIGPTDoubleHeadsModelOutput
+    :members:
+
 
 OpenAIGPTModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/model_doc/gpt2.rst

@@ -64,6 +64,9 @@ GPT2 specific outputs
 .. autoclass:: transformers.modeling_gpt2.GPT2DoubleHeadsModelOutput
     :members:
 
+.. autoclass:: transformers.modeling_tf_gpt2.TFGPT2DoubleHeadsModelOutput
+    :members:
+
 
 GPT2Model
 ~~~~~~~~~~~~~~~~~~~~~

docs/source/model_doc/mobilebert.rst

@@ -59,7 +59,10 @@ MobileBertTokenizerFast
 MobileBert specific outputs
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. autoclass:: transformers.modeling_mobilebert.MobileBertForPretrainingOutput
+.. autoclass:: transformers.modeling_mobilebert.MobileBertForPreTrainingOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_mobilebert.TFMobileBertForPreTrainingOutput
     :members:
 
 

docs/source/model_doc/transformerxl.rst

@@ -63,6 +63,12 @@ TransfoXL specific outputs
 .. autoclass:: transformers.modeling_transfo_xl.TransfoXLLMHeadModelOutput
     :members:
 
+.. autoclass:: transformers.modeling_tf_transfo_xl.TFTransfoXLModelOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_transfo_xl.TFTransfoXLLMHeadModelOutput
+    :members:
+
 
 TransfoXLModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/model_doc/xlnet.rst

@@ -74,6 +74,24 @@ XLNet specific outputs
 .. autoclass:: transformers.modeling_xlnet.XLNetForQuestionAnsweringOutput
     :members:
 
+.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetModelOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetLMHeadModelOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForSequenceClassificationOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForMultipleChoiceOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForTokenClassificationOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForQuestionAnsweringSimpleOutput
+    :members:
+
 
 XLNetModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

src/transformers/file_utils.py

@@ -190,7 +190,7 @@ def add_end_docstrings(*docstr):
     return docstring_decorator
 
 
-RETURN_INTRODUCTION = r"""
+PT_RETURN_INTRODUCTION = r"""
     Returns:
         :class:`~{full_output_type}` or :obj:`tuple(torch.FloatTensor)`:
         A :class:`~{full_output_type}` (if ``return_dict=True`` is passed or when ``config.return_dict=True``) or a
@@ -200,6 +200,16 @@ RETURN_INTRODUCTION = r"""
 """
 
 
+TF_RETURN_INTRODUCTION = r"""
+    Returns:
+        :class:`~{full_output_type}` or :obj:`tuple(tf.Tensor)`:
+        A :class:`~{full_output_type}` (if ``return_dict=True`` is passed or when ``config.return_dict=True``) or a
+        tuple of :obj:`tf.Tensor` comprising various elements depending on the configuration
+        (:class:`~transformers.{config_class}`) and inputs.
+
+"""
+
+
 def _get_indent(t):
     """Returns the indentation in the first line of t"""
     search = re.search(r"^(\s*)\S", t)
@@ -249,7 +259,8 @@ def _prepare_output_docstrings(output_type, config_class):
 
     # Add the return introduction
     full_output_type = f"{output_type.__module__}.{output_type.__name__}"
-    intro = RETURN_INTRODUCTION.format(full_output_type=full_output_type, config_class=config_class)
+    intro = TF_RETURN_INTRODUCTION if output_type.__name__.startswith("TF") else PT_RETURN_INTRODUCTION
+    intro = intro.format(full_output_type=full_output_type, config_class=config_class)
     return intro + docstrings
 
 

src/transformers/modeling_albert.py

@@ -407,9 +407,9 @@ class AlbertPreTrainedModel(PreTrainedModel):
 
 
 @dataclass
-class AlbertForPretrainingOutput(ModelOutput):
+class AlbertForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.AlbertForPretrainingModel`.
+    Output type of :class:`~transformers.AlbertForPreTrainingModel`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -643,7 +643,7 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
         return self.predictions.decoder
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
-    @replace_return_docstrings(output_type=AlbertForPretrainingOutput, config_class=_CONFIG_FOR_DOC)
+    @replace_return_docstrings(output_type=AlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids=None,
@@ -728,7 +728,7 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
             output = (prediction_scores, sop_scores) + outputs[2:]
             return ((total_loss,) + output) if total_loss is not None else output
 
-        return AlbertForPretrainingOutput(
+        return AlbertForPreTrainingOutput(
             loss=total_loss,
             prediction_logits=prediction_scores,
             sop_logits=sop_scores,

src/transformers/modeling_bert.py

@@ -586,9 +586,9 @@ class BertPreTrainedModel(PreTrainedModel):
 
 
 @dataclass
-class BertForPretrainingOutput(ModelOutput):
+class BertForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.BertForPretrainingModel`.
+    Output type of :class:`~transformers.BertForPreTrainingModel`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -837,7 +837,7 @@ class BertForPreTraining(BertPreTrainedModel):
         return self.cls.predictions.decoder
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @replace_return_docstrings(output_type=BertForPretrainingOutput, config_class=_CONFIG_FOR_DOC)
+    @replace_return_docstrings(output_type=BertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids=None,
@@ -918,7 +918,7 @@ class BertForPreTraining(BertPreTrainedModel):
             output = (prediction_scores, seq_relationship_score) + outputs[2:]
             return ((total_loss,) + output) if total_loss is not None else output
 
-        return BertForPretrainingOutput(
+        return BertForPreTrainingOutput(
             loss=total_loss,
             prediction_logits=prediction_scores,
             seq_relationship_logits=seq_relationship_score,

src/transformers/modeling_electra.py

@@ -188,9 +188,9 @@ class ElectraPreTrainedModel(BertPreTrainedModel):
 
 
 @dataclass
-class ElectraForPretrainingOutput(ModelOutput):
+class ElectraForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.ElectraForPretrainingModel`.
+    Output type of :class:`~transformers.ElectraForPreTrainingModel`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -496,7 +496,7 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
         self.init_weights()
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
-    @replace_return_docstrings(output_type=ElectraForPretrainingOutput, config_class=_CONFIG_FOR_DOC)
+    @replace_return_docstrings(output_type=ElectraForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids=None,
@@ -562,7 +562,7 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
             output = (logits,) + discriminator_hidden_states[1:]
             return ((loss,) + output) if loss is not None else output
 
-        return ElectraForPretrainingOutput(
+        return ElectraForPreTrainingOutput(
             loss=loss,
             logits=logits,
             hidden_states=discriminator_hidden_states.hidden_states,
@@ -850,7 +850,7 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
 @add_start_docstrings(
     """ELECTRA Model with a multiple choice classification head on top (a linear layer on top of
     the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
-    ELECTRA_INPUTS_DOCSTRING,
+    ELECTRA_START_DOCSTRING,
 )
 class ElectraForMultipleChoice(ElectraPreTrainedModel):
     def __init__(self, config):

src/transformers/modeling_mobilebert.py

@@ -685,9 +685,9 @@ class MobileBertPreTrainedModel(PreTrainedModel):
 
 
 @dataclass
-class MobileBertForPretrainingOutput(ModelOutput):
+class MobileBertForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.MobileBertForPretrainingModel`.
+    Output type of :class:`~transformers.MobileBertForPreTrainingModel`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -948,7 +948,7 @@ class MobileBertForPreTraining(MobileBertPreTrainedModel):
             self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING)
-    @replace_return_docstrings(output_type=MobileBertForPretrainingOutput, config_class=_CONFIG_FOR_DOC)
+    @replace_return_docstrings(output_type=MobileBertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def forward(
         self,
         input_ids=None,
@@ -1018,7 +1018,7 @@ class MobileBertForPreTraining(MobileBertPreTrainedModel):
             output = (prediction_scores, seq_relationship_score) + outputs[2:]
             return ((total_loss,) + output) if total_loss is not None else output
 
-        return MobileBertForPretrainingOutput(
+        return MobileBertForPreTrainingOutput(
             loss=total_loss,
             prediction_logits=prediction_scores,
             seq_relationship_logits=seq_relationship_score,

src/transformers/modeling_t5.py

@@ -973,7 +973,7 @@ class T5Model(T5PreTrainedModel):
                 output_hidden_states=output_hidden_states,
                 return_dict=return_dict,
             )
-        elif not return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
             encoder_outputs = BaseModelOutput(
                 last_hidden_state=encoder_outputs[0],
                 hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,

src/transformers/modeling_tf_albert.py

@@ -17,17 +17,30 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import Optional, Tuple
 
 import tensorflow as tf
 
 from .configuration_albert import AlbertConfig
 from .file_utils import (
     MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_callable,
+    replace_return_docstrings,
 )
 from .modeling_tf_bert import ACT2FN, TFBertSelfAttention
+from .modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPooling,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMaskedLanguageModelingLoss,
     TFMultipleChoiceLoss,
@@ -44,6 +57,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "AlbertConfig"
 _TOKENIZER_FOR_DOC = "AlbertTokenizer"
 
 TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -414,12 +428,19 @@ class TFAlbertTransformer(tf.keras.layers.Layer):
             for i in range(config.num_hidden_groups)
         ]
 
-    def call(self, hidden_states, attention_mask, head_mask, output_attentions, output_hidden_states, training=False):
+    def call(
+        self,
+        hidden_states,
+        attention_mask,
+        head_mask,
+        output_attentions,
+        output_hidden_states,
+        return_dict,
+        training=False,
+    ):
         hidden_states = self.embedding_hidden_mapping_in(hidden_states)
-        all_attentions = ()
+        all_attentions = () if output_attentions else None
-
+        all_hidden_states = (hidden_states,) if output_hidden_states else None
-        if output_hidden_states:
-            all_hidden_states = (hidden_states,)
 
         for i in range(self.config.num_hidden_layers):
             # Number of layers in a hidden group
@@ -444,14 +465,11 @@ class TFAlbertTransformer(tf.keras.layers.Layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states,)
+        if not return_dict:
-        if output_hidden_states:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
-            outputs = outputs + (all_hidden_states,)
+        return TFBaseModelOutput(
-        if output_attentions:
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
-            outputs = outputs + (all_attentions,)
+        )
-
-        # last-layer hidden state, (all hidden states), (all attentions)
-        return outputs
 
 
 class TFAlbertPreTrainedModel(TFPreTrainedModel):
@@ -506,6 +524,7 @@ class TFAlbertMainLayer(tf.keras.layers.Layer):
         self.num_hidden_layers = config.num_hidden_layers
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.return_dict = config.use_return_dict
 
         self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
         self.encoder = TFAlbertTransformer(config, name="encoder")
@@ -543,6 +562,7 @@ class TFAlbertMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -554,7 +574,8 @@ class TFAlbertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -564,12 +585,14 @@ class TFAlbertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -619,16 +642,52 @@ class TFAlbertMainLayer(tf.keras.layers.Layer):
             head_mask,
             output_attentions,
             output_hidden_states,
+            return_dict,
             training=training,
         )
 
         sequence_output = encoder_outputs[0]
         pooled_output = self.pooler(sequence_output[:, 0])
 
-        # add hidden_states and attentions if they are here
+        if not return_dict:
-        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]
+            return (sequence_output, pooled_output,) + encoder_outputs[1:]
-        # sequence_output, pooled_output, (hidden_states), (attentions)
+
-        return outputs
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@dataclass
+class TFAlbertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFAlbertForPreTrainingModel`.
+
+    Args:
+        prediction_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        sop_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False
+            continuation before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    prediction_logits: tf.Tensor = None
+    sop_logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
 
 
 ALBERT_START_DOCSTRING = r"""
@@ -707,6 +766,11 @@ ALBERT_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -720,32 +784,13 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
         self.albert = TFAlbertMainLayer(config, name="albert")
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="albert-base-v2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="albert-base-v2",
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`tf.Tensor` of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during Albert pretraining. This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.albert(inputs, **kwargs)
         return outputs
 
@@ -768,25 +813,10 @@ class TFAlbertForPreTraining(TFAlbertPreTrainedModel):
         return self.albert.embeddings
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @replace_return_docstrings(output_type=TFAlbertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def call(self, inputs, **kwargs):
         r"""
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        sop_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, 2)`):
-            Prediction scores of the sentence order prediction (classification) head (scores of True/False continuation before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
         import tensorflow as tf
@@ -797,13 +827,22 @@ class TFAlbertForPreTraining(TFAlbertPreTrainedModel):
         outputs = model(input_ids)
         prediction_scores, sop_scores = outputs[:2]
         """
-
+        return_dict = kwargs.get("return_dict")
+        return_dict = return_dict if return_dict is not None else self.albert.return_dict
         outputs = self.albert(inputs, **kwargs)
         sequence_output, pooled_output = outputs[:2]
         prediction_scores = self.predictions(sequence_output)
         sop_scores = self.sop_classifier(pooled_output, training=kwargs.get("training", False))
-        outputs = (prediction_scores, sop_scores) + outputs[2:]
+
-        return outputs
+        if not return_dict:
+            return (prediction_scores, sop_scores) + outputs[2:]
+
+        return TFAlbertForPreTrainingOutput(
+            prediction_logits=prediction_scores,
+            sop_logits=sop_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
 
 
 class TFAlbertSOPHead(tf.keras.layers.Layer):
@@ -833,7 +872,12 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss)
         return self.albert.embeddings
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="albert-base-v2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="albert-base-v2",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -844,6 +888,7 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss)
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -853,27 +898,12 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss)
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.albert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -886,20 +916,22 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel, TFMaskedLanguageModelingLoss)
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
         sequence_output = outputs[0]
         prediction_scores = self.predictions(sequence_output, training=training)
 
-        # Add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, prediction_scores)
-        outputs = (prediction_scores,) + outputs[2:]
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, prediction_scores)
+            output = (prediction_scores,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # prediction_scores, (hidden_states), (attentions)
+        return TFMaskedLMOutput(
+            loss=loss, logits=prediction_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -919,7 +951,12 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
         )
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="albert-base-v2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="albert-base-v2",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -930,6 +967,7 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -939,27 +977,12 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
             Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`)
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.albert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -972,6 +995,7 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -980,13 +1004,15 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1006,7 +1032,12 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
         )
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="albert-base-v2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="albert-base-v2",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1017,6 +1048,7 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1024,27 +1056,12 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.albert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1057,6 +1074,7 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1065,13 +1083,15 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1089,7 +1109,12 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
         )
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="albert-base-v2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="albert-base-v2",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1100,6 +1125,7 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -1113,30 +1139,13 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.albert.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[8] if len(inputs) > 8 else start_positions
+            start_positions = inputs[9] if len(inputs) > 9 else start_positions
-            end_positions = inputs[9] if len(inputs) > 9 else end_positions
+            end_positions = inputs[10] if len(inputs) > 10 else end_positions
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -1150,6 +1159,7 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1160,15 +1170,23 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + outputs[2:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1196,7 +1214,12 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="albert-base-v2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="albert-base-v2",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -1207,6 +1230,7 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1215,24 +1239,6 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -1243,8 +1249,9 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            labels = inputs[8] if len(inputs) > 8 else labels
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
-            assert len(inputs) <= 9, "Too many inputs."
+            labels = inputs[9] if len(inputs) > 9 else labels
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -1254,10 +1261,12 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_attentions)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 9, "Too many inputs."
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.albert.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -1280,6 +1289,7 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1289,10 +1299,12 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
         logits = self.classifier(pooled_output)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
 
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, reshaped_logits)
+            output = (reshaped_logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        return TFMultipleChoiceModelOutput(
+            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )

src/transformers/modeling_tf_bert.py

@@ -17,6 +17,8 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import Optional, Tuple
 
 import numpy as np
 import tensorflow as tf
@@ -24,9 +26,22 @@ import tensorflow as tf
 from .configuration_bert import BertConfig
 from .file_utils import (
     MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_callable,
+    replace_return_docstrings,
+)
+from .modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPooling,
+    TFCausalLMOutput,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFNextSentencePredictorOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
 )
 from .modeling_tf_utils import (
     TFCausalLanguageModelingLoss,
@@ -45,6 +60,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "BertConfig"
 _TOKENIZER_FOR_DOC = "BertTokenizer"
 
 TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -389,9 +405,18 @@ class TFBertEncoder(tf.keras.layers.Layer):
         super().__init__(**kwargs)
         self.layer = [TFBertLayer(config, name="layer_._{}".format(i)) for i in range(config.num_hidden_layers)]
 
-    def call(self, hidden_states, attention_mask, head_mask, output_attentions, output_hidden_states, training=False):
+    def call(
-        all_hidden_states = ()
+        self,
-        all_attentions = ()
+        hidden_states,
+        attention_mask,
+        head_mask,
+        output_attentions,
+        output_hidden_states,
+        return_dict,
+        training=False,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
 
         for i, layer_module in enumerate(self.layer):
             if output_hidden_states:
@@ -409,15 +434,11 @@ class TFBertEncoder(tf.keras.layers.Layer):
         if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states,)
+        if not return_dict:
-
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
-        if output_hidden_states:
+        return TFBaseModelOutput(
-            outputs = outputs + (all_hidden_states,)
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
-
+        )
-        if output_attentions:
-            outputs = outputs + (all_attentions,)
-
-        return outputs  # outputs, (hidden states), (attentions)
 
 
 class TFBertPooler(tf.keras.layers.Layer):
@@ -517,6 +538,7 @@ class TFBertMainLayer(tf.keras.layers.Layer):
         self.initializer_range = config.initializer_range
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.return_dict = config.use_return_dict
         self.embeddings = TFBertEmbeddings(config, name="embeddings")
         self.encoder = TFBertEncoder(config, name="encoder")
         self.pooler = TFBertPooler(config, name="pooler")
@@ -545,6 +567,7 @@ class TFBertMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -556,7 +579,8 @@ class TFBertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -566,12 +590,14 @@ class TFBertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -621,16 +647,22 @@ class TFBertMainLayer(tf.keras.layers.Layer):
             head_mask,
             output_attentions,
             output_hidden_states,
+            return_dict,
             training=training,
         )
 
         sequence_output = encoder_outputs[0]
         pooled_output = self.pooler(sequence_output)
-        outputs = (sequence_output, pooled_output,) + encoder_outputs[
-            1:
-        ]  # add hidden_states and attentions if they are here
 
-        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
+        if not return_dict:
+            return (sequence_output, pooled_output,) + encoder_outputs[1:]
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
 
 
 class TFBertPreTrainedModel(TFPreTrainedModel):
@@ -642,6 +674,36 @@ class TFBertPreTrainedModel(TFPreTrainedModel):
     base_model_prefix = "bert"
 
 
+@dataclass
+class TFBertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFBertForPreTrainingModel`.
+
+    Args:
+        prediction_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        seq_relationship_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False
+            continuation before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    prediction_logits: tf.Tensor = None
+    seq_relationship_logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 BERT_START_DOCSTRING = r"""
     This model is a `tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ sub-class.
     Use it as a regular TF 2.0 Keras Model and
@@ -712,6 +774,11 @@ BERT_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -725,32 +792,13 @@ class TFBertModel(TFBertPreTrainedModel):
         self.bert = TFBertMainLayer(config, name="bert")
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`tf.Tensor` of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during Bert pretraining. This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.bert(inputs, **kwargs)
         return outputs
 
@@ -772,25 +820,10 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
         return self.bert.embeddings
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @replace_return_docstrings(output_type=TFBertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def call(self, inputs, **kwargs):
         r"""
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        seq_relationship_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, 2)`):
-            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -804,17 +837,23 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
         prediction_scores, seq_relationship_scores = outputs[:2]
 
         """
+        return_dict = kwargs.get("return_dict")
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         outputs = self.bert(inputs, **kwargs)
 
         sequence_output, pooled_output = outputs[:2]
         prediction_scores = self.mlm(sequence_output, training=kwargs.get("training", False))
         seq_relationship_score = self.nsp(pooled_output)
 
-        outputs = (prediction_scores, seq_relationship_score,) + outputs[
+        if not return_dict:
-            2:
+            return (prediction_scores, seq_relationship_score) + outputs[2:]
-        ]  # add hidden states and attention if they are here
 
-        return outputs  # prediction_scores, seq_relationship_score, (hidden_states), (attentions)
+        return TFBertForPreTrainingOutput(
+            prediction_logits=prediction_scores,
+            seq_relationship_logits=seq_relationship_score,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings("""Bert Model with a `language modeling` head on top. """, BERT_START_DOCSTRING)
@@ -832,7 +871,12 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
         return self.bert.embeddings
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -843,6 +887,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -852,27 +897,12 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -885,19 +915,22 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
         sequence_output = outputs[0]
         prediction_scores = self.mlm(sequence_output, training=training)
 
-        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, prediction_scores)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, prediction_scores)
+            output = (prediction_scores,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), prediction_scores, (hidden_states), (attentions)
+        return TFMaskedLMOutput(
+            loss=loss, logits=prediction_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
@@ -911,7 +944,12 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
     def get_output_embeddings(self):
         return self.bert.embeddings
 
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFCausalLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -922,6 +960,7 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -929,27 +968,12 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the cross entropy classification loss.
             Indices should be in ``[0, ..., config.vocab_size - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -962,21 +986,27 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
         sequence_output = outputs[0]
         logits = self.mlm(sequence_output, training=training)
 
-        outputs = (logits,) + outputs[2:]  # Add hidden states and attention if they are here
+        loss = None
         if labels is not None:
             # shift labels to the left and cut last logit token
             logits = logits[:, :-1]
             labels = labels[:, 1:]
             loss = self.compute_loss(labels, logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # prediction_scores, (hidden_states), (attentions)
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFCausalLMOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -990,23 +1020,10 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
         self.nsp = TFBertNSPHead(config, name="nsp___cls")
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @replace_return_docstrings(output_type=TFNextSentencePredictorOutput, config_class=_CONFIG_FOR_DOC)
     def call(self, inputs, **kwargs):
         r"""
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        seq_relationship_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, 2)`)
-            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -1023,14 +1040,19 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
         logits = model(encoding['input_ids'], token_type_ids=encoding['token_type_ids'])[0]
         assert logits[0][0] < logits[0][1] # the next sentence was random
         """
+        return_dict = kwargs.get("return_dict")
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         outputs = self.bert(inputs, **kwargs)
 
         pooled_output = outputs[1]
         seq_relationship_score = self.nsp(pooled_output)
 
-        outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
+        if not return_dict:
+            return (seq_relationship_score,) + outputs[2:]
 
-        return outputs  # seq_relationship_score, (hidden_states), (attentions)
+        return TFNextSentencePredictorOutput(
+            logits=seq_relationship_score, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1050,7 +1072,12 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
         )
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1061,6 +1088,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1070,27 +1098,12 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
             Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
             If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
             If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1103,6 +1116,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1111,13 +1125,15 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1145,7 +1161,12 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -1156,6 +1177,7 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1164,24 +1186,6 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -1192,8 +1196,9 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            labels = inputs[8] if len(inputs) > 8 else labels
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
-            assert len(inputs) <= 9, "Too many inputs."
+            labels = inputs[9] if len(inputs) > 9 else labels
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -1203,10 +1208,12 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 9, "Too many inputs."
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -1233,19 +1240,23 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
             flat_inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         pooled_output = outputs[1]
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
-            loss = self.compute_loss(labels, reshaped_logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1265,7 +1276,12 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
         )
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1276,6 +1292,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1283,27 +1300,12 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1316,6 +1318,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1324,13 +1327,15 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1349,7 +1354,12 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
         )
 
     @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="bert-base-cased",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1360,6 +1370,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -1373,30 +1384,13 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.bert.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[8] if len(inputs) > 8 else start_positions
+            start_positions = inputs[9] if len(inputs) > 9 else start_positions
-            end_positions = inputs[9] if len(inputs) > 9 else end_positions
+            end_positions = inputs[10] if len(inputs) > 10 else end_positions
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -1410,6 +1404,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1420,12 +1415,20 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + outputs[2:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/modeling_tf_camembert.py

@@ -62,8 +62,6 @@ CAMEMBERT_START_DOCSTRING = r"""
         config (:class:`~transformers.CamembertConfig`): Model configuration class with all the parameters of the
             model. Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
-        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
-            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 

src/transformers/modeling_tf_ctrl.py

@@ -23,6 +23,7 @@ import tensorflow as tf
 
 from .configuration_ctrl import CTRLConfig
 from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
+from .modeling_tf_outputs import TFBaseModelOutputWithPast, TFCausalLMOutputWithPast
 from .modeling_tf_utils import (
     TFCausalLanguageModelingLoss,
     TFPreTrainedModel,
@@ -35,7 +36,8 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
-_TOKENIZER_FOR_DOC = "CtrlTokenizer"
+_CONFIG_FOR_DOC = "CTRLConfig"
+_TOKENIZER_FOR_DOC = "CTRLTokenizer"
 
 TF_CTRL_PRETRAINED_MODEL_ARCHIVE_LIST = [
     "ctrl"
@@ -207,6 +209,7 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
         self.use_cache = config.use_cache
+        self.return_dict = config.use_return_dict
 
         self.d_model_size = config.n_embd
         self.num_layers = config.n_layer
@@ -260,6 +263,7 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
 
@@ -274,7 +278,8 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
             use_cache = inputs[7] if len(inputs) > 7 else use_cache
             output_attentions = inputs[8] if len(inputs) > 8 else output_attentions
             output_hidden_states = inputs[9] if len(inputs) > 9 else output_hidden_states
-            assert len(inputs) <= 10, "Too many inputs."
+            return_dict = inputs[10] if len(inputs) > 10 else return_dict
+            assert len(inputs) <= 11, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             past = inputs.get("past", past)
@@ -286,13 +291,15 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 10, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 11, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
         use_cache = use_cache if use_cache is not None else self.use_cache
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         # If using past key value states, only the last tokens
         # should be given as an input
@@ -374,9 +381,9 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
         hidden_states = self.dropout(hidden_states, training=training)
 
         output_shape = input_shape + [shape_list(hidden_states)[-1]]
-        presents = ()
+        presents = () if use_cache else None
-        all_hidden_states = ()
+        all_hidden_states = () if output_hidden_states else None
-        all_attentions = []
+        all_attentions = () if output_attentions else None
         for i, (h, layer_past) in enumerate(zip(self.h, past)):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
@@ -396,24 +403,27 @@ class TFCTRLMainLayer(tf.keras.layers.Layer):
                 presents = presents + (present,)
 
             if output_attentions:
-                all_attentions.append(outputs[2])
+                all_attentions = all_attentions + (outputs[2],)
 
         hidden_states = self.layernorm(hidden_states)
         hidden_states = tf.reshape(hidden_states, output_shape)
         if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states,)
-        if use_cache:
-            outputs = outputs + (presents,)
-        if output_hidden_states:
-            outputs = outputs + (all_hidden_states,)
         if output_attentions:
             # let the number of heads free (-1) so we can extract attention even after head pruning
             attention_output_shape = input_shape[:-1] + [-1] + shape_list(all_attentions[0])[-2:]
             all_attentions = tuple(tf.reshape(t, attention_output_shape) for t in all_attentions)
-            outputs = outputs + (all_attentions,)
+
-        return outputs
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
 
 
 class TFCTRLPreTrainedModel(TFPreTrainedModel):
@@ -503,6 +513,11 @@ CTRL_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -516,29 +531,13 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
         self.transformer = TFCTRLMainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(CTRL_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="ctrl")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="ctrl",
+        output_type=TFBaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.CTRLConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the last layer of the model.
-        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -585,7 +584,12 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
         return {"inputs": inputs, "past": past, "use_cache": kwargs["use_cache"]}
 
     @add_start_docstrings_to_callable(CTRL_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="ctrl")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="ctrl",
+        output_type=TFCausalLMOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -598,6 +602,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -605,31 +610,12 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the cross entropy classification loss.
             Indices should be in ``[0, ..., config.vocab_size - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.CTRLConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[10] if len(inputs) > 10 else labels
+            labels = inputs[11] if len(inputs) > 11 else labels
-            if len(inputs) > 10:
+            if len(inputs) > 11:
-                inputs = inputs[:10]
+                inputs = inputs[:11]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -644,6 +630,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -651,12 +638,21 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel, TFCausalLanguageModelingLoss):
 
         logits = self.lm_head(hidden_states)
 
-        outputs = (logits,) + transformer_outputs[1:]
+        loss = None
         if labels is not None:
             # shift labels to the left and cut last logit token
             logits = logits[:, :-1]
             labels = labels[:, 1:]
             loss = self.compute_loss(labels, logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # lm_logits, presents, (all hidden_states), (attentions)
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

src/transformers/modeling_tf_distilbert.py

@@ -29,6 +29,14 @@ from .file_utils import (
     add_start_docstrings,
     add_start_docstrings_to_callable,
 )
+from .modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMaskedLanguageModelingLoss,
     TFMultipleChoiceLoss,
@@ -46,6 +54,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "DistilBertConfig"
 _TOKENIZER_FOR_DOC = "DistilBertTokenizer"
 
 TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -359,7 +368,7 @@ class TFTransformer(tf.keras.layers.Layer):
 
         self.layer = [TFTransformerBlock(config, name="layer_._{}".format(i)) for i in range(config.n_layers)]
 
-    def call(self, x, attn_mask, head_mask, output_attentions, output_hidden_states, training=False):
+    def call(self, x, attn_mask, head_mask, output_attentions, output_hidden_states, return_dict, training=False):
         """
         Parameters
         ----------
@@ -379,8 +388,8 @@ class TFTransformer(tf.keras.layers.Layer):
             Tuple of length n_layers with the attention weights from each layer
             Optional: only if output_attentions=True
         """
-        all_hidden_states = ()
+        all_hidden_states = () if output_hidden_states else None
-        all_attentions = ()
+        all_attentions = () if output_attentions else None
 
         hidden_state = x
         for i, layer_module in enumerate(self.layer):
@@ -401,12 +410,11 @@ class TFTransformer(tf.keras.layers.Layer):
         if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_state,)
 
-        outputs = (hidden_state,)
+        if not return_dict:
-        if output_hidden_states:
+            return tuple(v for v in [hidden_state, all_hidden_states, all_attentions] if v is not None)
-            outputs = outputs + (all_hidden_states,)
+        return TFBaseModelOutput(
-        if output_attentions:
+            last_hidden_state=hidden_state, hidden_states=all_hidden_states, attentions=all_attentions
-            outputs = outputs + (all_attentions,)
+        )
-        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
 
 
 @keras_serializable
@@ -418,6 +426,7 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
         self.num_hidden_layers = config.num_hidden_layers
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.return_dict = config.use_return_dict
 
         self.embeddings = TFEmbeddings(config, name="embeddings")  # Embeddings
         self.transformer = TFTransformer(config, name="transformer")  # Encoder
@@ -440,6 +449,7 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -449,7 +459,8 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
             output_attentions = inputs[4] if len(inputs) > 4 else output_attentions
             output_hidden_states = inputs[5] if len(inputs) > 5 else output_hidden_states
-            assert len(inputs) <= 6, "Too many inputs."
+            return_dict = inputs[6] if len(inputs) > 6 else return_dict
+            assert len(inputs) <= 7, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -457,12 +468,14 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 6, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 7, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -491,7 +504,13 @@ class TFDistilBertMainLayer(tf.keras.layers.Layer):
 
         embedding_output = self.embeddings(input_ids, inputs_embeds=inputs_embeds)  # (bs, seq_length, dim)
         tfmr_output = self.transformer(
-            embedding_output, attention_mask, head_mask, output_attentions, output_hidden_states, training=training
+            embedding_output,
+            attention_mask,
+            head_mask,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            training=training,
         )
 
         return tfmr_output  # last-layer hidden-state, (all hidden_states), (all attentions)
@@ -564,9 +583,13 @@ DISTILBERT_INPUTS_DOCSTRING = r"""
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
-
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -580,25 +603,13 @@ class TFDistilBertModel(TFDistilBertPreTrainedModel):
         self.distilbert = TFDistilBertMainLayer(config, name="distilbert")  # Embeddings
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="distilbert-base-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="distilbert-base-uncased",
+        output_type=TFBaseModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers,DistilBertConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.distilbert(inputs, **kwargs)
         return outputs
 
@@ -642,7 +653,12 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel
         return self.vocab_projector.input_embeddings
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="distilbert-base-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="distilbert-base-uncased",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -651,6 +667,7 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -660,27 +677,12 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers,DistilBertConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.distilbert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[6] if len(inputs) > 6 else labels
+            labels = inputs[7] if len(inputs) > 7 else labels
-            if len(inputs) > 6:
+            if len(inputs) > 7:
-                inputs = inputs[:6]
+                inputs = inputs[:7]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -691,6 +693,7 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -700,13 +703,18 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel, TFMaskedLanguageModel
         prediction_logits = self.vocab_layer_norm(prediction_logits)  # (bs, seq_length, dim)
         prediction_logits = self.vocab_projector(prediction_logits)
 
-        outputs = (prediction_logits,) + distilbert_output[1:]
+        loss = None if labels is None else self.compute_loss(labels, prediction_logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, prediction_logits)
+            output = (prediction_logits,) + distilbert_output[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # logits, (hidden_states), (attentions)
+        return TFMaskedLMOutput(
+            loss=loss,
+            logits=prediction_logits,
+            hidden_states=distilbert_output.hidden_states,
+            attentions=distilbert_output.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -732,7 +740,12 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
         self.dropout = tf.keras.layers.Dropout(config.seq_classif_dropout)
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="distilbert-base-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="distilbert-base-uncased",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -741,6 +754,7 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -750,27 +764,12 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
             Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers,DistilBertConfig`) and inputs:
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.distilbert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[6] if len(inputs) > 6 else labels
+            labels = inputs[7] if len(inputs) > 7 else labels
-            if len(inputs) > 6:
+            if len(inputs) > 7:
-                inputs = inputs[:6]
+                inputs = inputs[:7]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -781,6 +780,7 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -790,13 +790,18 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
         pooled_output = self.dropout(pooled_output, training=training)  # (bs, dim)
         logits = self.classifier(pooled_output)  # (bs, dim)
 
-        outputs = (logits,) + distilbert_output[1:]
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + distilbert_output[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=distilbert_output.hidden_states,
+            attentions=distilbert_output.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -816,7 +821,12 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla
         )
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="distilbert-base-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="distilbert-base-uncased",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -825,6 +835,7 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -832,27 +843,12 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers,DistilBertConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.distilbert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[6] if len(inputs) > 6 else labels
+            labels = inputs[7] if len(inputs) > 7 else labels
-            if len(inputs) > 6:
+            if len(inputs) > 7:
-                inputs = inputs[:6]
+                inputs = inputs[:7]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -863,6 +859,7 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -871,13 +868,15 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + outputs[1:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -911,7 +910,12 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="distilbert-base-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="distilbert-base-uncased",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -920,6 +924,7 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -928,24 +933,6 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -954,8 +941,9 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
             inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
             output_attentions = inputs[4] if len(inputs) > 4 else output_attentions
             output_hidden_states = inputs[5] if len(inputs) > 5 else output_hidden_states
-            labels = inputs[6] if len(inputs) > 6 else labels
+            return_dict = inputs[6] if len(inputs) > 6 else return_dict
-            assert len(inputs) <= 7, "Too many inputs."
+            labels = inputs[7] if len(inputs) > 7 else labels
+            assert len(inputs) <= 8, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -963,10 +951,12 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 7, "Too many inputs."
+            assert len(inputs) <= 8, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.distilbert.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -989,6 +979,7 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
             flat_inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         hidden_state = distilbert_output[0]  # (bs, seq_len, dim)
@@ -997,13 +988,19 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
         pooled_output = self.dropout(pooled_output, training=training)  # (bs, dim)
         logits = self.classifier(pooled_output)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + distilbert_output[1:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
-            loss = self.compute_loss(labels, reshaped_logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (reshaped_logits,) + distilbert_output[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=distilbert_output.hidden_states,
+            attentions=distilbert_output.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1023,7 +1020,12 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
         self.dropout = tf.keras.layers.Dropout(config.qa_dropout)
 
     @add_start_docstrings_to_callable(DISTILBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="distilbert-base-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="distilbert-base-uncased",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1032,6 +1034,7 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -1045,30 +1048,13 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers,DistilBertConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.distilbert.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[6] if len(inputs) > 6 else start_positions
+            start_positions = inputs[7] if len(inputs) > 7 else start_positions
-            end_positions = inputs[7] if len(inputs) > 7 else end_positions
+            end_positions = inputs[8] if len(inputs) > 8 else end_positions
-            if len(inputs) > 6:
+            if len(inputs) > 7:
-                inputs = inputs[:6]
+                inputs = inputs[:7]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -1080,6 +1066,7 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1090,12 +1077,20 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + distilbert_output[1:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + distilbert_output[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=distilbert_output.hidden_states,
+            attentions=distilbert_output.attentions,
+        )

src/transformers/modeling_tf_electra.py

@@ -1,4 +1,6 @@
 import logging
+from dataclasses import dataclass
+from typing import Optional, Tuple
 
 import tensorflow as tf
 
@@ -6,11 +8,21 @@ from transformers import ElectraConfig
 
 from .file_utils import (
     MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_callable,
+    replace_return_docstrings,
 )
 from .modeling_tf_bert import ACT2FN, TFBertEncoder, TFBertPreTrainedModel
+from .modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMaskedLanguageModelingLoss,
     TFMultipleChoiceLoss,
@@ -27,8 +39,8 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
-_TOKENIZER_FOR_DOC = "ElectraTokenizer"
 _CONFIG_FOR_DOC = "ElectraConfig"
+_TOKENIZER_FOR_DOC = "ElectraTokenizer"
 
 TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = [
     "google/electra-small-generator",
@@ -254,6 +266,7 @@ class TFElectraMainLayer(TFElectraPreTrainedModel):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -265,7 +278,8 @@ class TFElectraMainLayer(TFElectraPreTrainedModel):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -275,7 +289,8 @@ class TFElectraMainLayer(TFElectraPreTrainedModel):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
@@ -283,6 +298,7 @@ class TFElectraMainLayer(TFElectraPreTrainedModel):
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -312,12 +328,41 @@ class TFElectraMainLayer(TFElectraPreTrainedModel):
             head_mask,
             output_attentions,
             output_hidden_states,
+            return_dict,
             training=training,
         )
 
         return hidden_states
 
 
+@dataclass
+class TFElectraForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFElectraForPreTrainingModel`.
+
+    Args:
+        loss (`optional`, returned when ``labels`` is provided, ``tf.Tensor`` of shape :obj:`(1,)`):
+            Total loss of the ELECTRA objective.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+            Prediction scores of the head (scores for each token before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 ELECTRA_START_DOCSTRING = r"""
     This model is a `tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ sub-class.
     Use it as a regular TF 2.0 Keras Model and
@@ -380,9 +425,13 @@ ELECTRA_INPUTS_DOCSTRING = r"""
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
-
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -400,25 +449,13 @@ class TFElectraModel(TFElectraPreTrainedModel):
         self.electra = TFElectraMainLayer(config, name="electra")
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/electra-small-discriminator")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/electra-small-discriminator",
+        output_type=TFBaseModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.electra(inputs, **kwargs)
         return outputs
 
@@ -439,6 +476,7 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel):
         self.discriminator_predictions = TFElectraDiscriminatorPredictions(config, name="discriminator_predictions")
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFElectraForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def call(
         self,
         input_ids,
@@ -449,24 +487,11 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         r"""
     Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Prediction scores of the head (scores for each token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -479,6 +504,7 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel):
         outputs = model(input_ids)
         scores = outputs[0]
         """
+        return_dict = return_dict if return_dict is not None else self.electra.config.return_dict
 
         discriminator_hidden_states = self.electra(
             input_ids,
@@ -489,14 +515,20 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel):
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         discriminator_sequence_output = discriminator_hidden_states[0]
         logits = self.discriminator_predictions(discriminator_sequence_output)
-        output = (logits,)
-        output += discriminator_hidden_states[1:]
 
-        return output  # (loss), scores, (hidden_states), (attentions)
+        if not return_dict:
+            return (logits,) + discriminator_hidden_states[1:]
+
+        return TFElectraForPreTrainingOutput(
+            logits=logits,
+            hidden_states=discriminator_hidden_states.hidden_states,
+            attentions=discriminator_hidden_states.attentions,
+        )
 
 
 class TFElectraMaskedLMHead(tf.keras.layers.Layer):
@@ -539,7 +571,12 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLos
         return self.generator_lm_head
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/electra-small-generator")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/electra-small-generator",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         input_ids,
@@ -550,6 +587,7 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLos
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -559,27 +597,12 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLos
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.electra.config.return_dict
         if isinstance(input_ids, (tuple, list)):
-            labels = input_ids[8] if len(input_ids) > 8 else labels
+            labels = input_ids[9] if len(input_ids) > 9 else labels
-            if len(input_ids) > 8:
+            if len(input_ids) > 9:
-                input_ids = input_ids[:8]
+                input_ids = input_ids[:9]
         elif isinstance(input_ids, (dict, BatchEncoding)):
             labels = input_ids.pop("labels", labels)
 
@@ -592,19 +615,25 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLos
             inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         generator_sequence_output = generator_hidden_states[0]
         prediction_scores = self.generator_predictions(generator_sequence_output, training=training)
         prediction_scores = self.generator_lm_head(prediction_scores, training=training)
-        output = (prediction_scores,)
-        output += generator_hidden_states[1:]
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, prediction_scores)
-            loss = self.compute_loss(labels, prediction_scores)
-            output = (loss,) + output
 
-        return output  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+        if not return_dict:
+            output = (prediction_scores,) + generator_hidden_states[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMaskedLMOutput(
+            loss=loss,
+            logits=prediction_scores,
+            hidden_states=generator_hidden_states.hidden_states,
+            attentions=generator_hidden_states.attentions,
+        )
 
 
 class TFElectraClassificationHead(tf.keras.layers.Layer):
@@ -647,6 +676,7 @@ class TFElectraForSequenceClassification(TFElectraPreTrainedModel, TFSequenceCla
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
         checkpoint="google/electra-small-discriminator",
+        output_type=TFSequenceClassifierOutput,
         config_class=_CONFIG_FOR_DOC,
     )
     def call(
@@ -659,23 +689,25 @@ class TFElectraForSequenceClassification(TFElectraPreTrainedModel, TFSequenceCla
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
         r"""
-    Returns:
+        labels (:obj:`tf.Tensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
+            Labels for computing the sequence classification/regression loss.
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`)
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when :obj:`config.output_hidden_states=True`):
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
         """
+        return_dict = return_dict if return_dict is not None else self.electra.config.return_dict
+        if isinstance(input_ids, (tuple, list)):
+            labels = input_ids[9] if len(input_ids) > 9 else labels
+            if len(input_ids) > 9:
+                input_ids = input_ids[:9]
+        elif isinstance(input_ids, (dict, BatchEncoding)):
+            labels = input_ids.pop("labels", labels)
+
         outputs = self.electra(
             input_ids,
             attention_mask,
@@ -685,16 +717,20 @@ class TFElectraForSequenceClassification(TFElectraPreTrainedModel, TFSequenceCla
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         logits = self.classifier(outputs[0])
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, logits)
-            loss = self.compute_loss(labels, logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -724,7 +760,12 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/electra-small-discriminator")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/electra-small-discriminator",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -735,6 +776,7 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -743,24 +785,6 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -771,8 +795,9 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            labels = inputs[8] if len(inputs) > 8 else labels
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
-            assert len(inputs) <= 9, "Too many inputs."
+            labels = inputs[9] if len(inputs) > 9 else labels
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -782,10 +807,12 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 9, "Too many inputs."
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.electra.config.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -812,18 +839,22 @@ class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss)
             flat_inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         logits = self.sequence_summary(outputs[0])
         logits = self.classifier(logits)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
-            loss = self.compute_loss(labels, reshaped_logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -843,7 +874,12 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
         )
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/electra-small-discriminator")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/electra-small-discriminator",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -854,6 +890,7 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -861,27 +898,12 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.ElectraConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.electra.config.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -894,19 +916,25 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         discriminator_sequence_output = discriminator_hidden_states[0]
         discriminator_sequence_output = self.dropout(discriminator_sequence_output)
         logits = self.classifier(discriminator_sequence_output)
 
-        outputs = (logits,) + discriminator_hidden_states[1:]
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + discriminator_hidden_states[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), scores, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=discriminator_hidden_states.hidden_states,
+            attentions=discriminator_hidden_states.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -925,7 +953,12 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
         )
 
     @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/electra-small-discriminator")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/electra-small-discriminator",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -936,6 +969,7 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -949,30 +983,13 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.electra.config.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[8] if len(inputs) > 8 else start_positions
+            start_positions = inputs[9] if len(inputs) > 9 else start_positions
-            end_positions = inputs[9] if len(inputs) > 9 else end_positions
+            end_positions = inputs[10] if len(inputs) > 10 else end_positions
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -986,6 +1003,7 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         discriminator_sequence_output = discriminator_hidden_states[0]
@@ -995,12 +1013,20 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + discriminator_hidden_states[1:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits,) + discriminator_hidden_states[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=discriminator_hidden_states.hidden_states,
+            attentions=discriminator_hidden_states.attentions,
+        )

src/transformers/modeling_tf_flaubert.py

@@ -22,6 +22,7 @@ import tensorflow as tf
 
 from .configuration_flaubert import FlaubertConfig
 from .file_utils import add_start_docstrings
+from .modeling_tf_outputs import TFBaseModelOutput
 from .modeling_tf_utils import keras_serializable, shape_list
 from .modeling_tf_xlm import (
     TFXLMForMultipleChoice,
@@ -103,6 +104,11 @@ FLAUBERT_INPUTS_DOCSTRING = r"""
             than the model's internal embedding lookup matrix.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -126,6 +132,7 @@ class TFFlaubertMainLayer(TFXLMMainLayer):
         self.pre_norm = getattr(config, "pre_norm", False)
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.return_dict = config.use_return_dict
 
     def call(
         self,
@@ -140,6 +147,7 @@ class TFFlaubertMainLayer(TFXLMMainLayer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         # removed: src_enc=None, src_len=None
@@ -155,7 +163,8 @@ class TFFlaubertMainLayer(TFXLMMainLayer):
             inputs_embeds = inputs[8] if len(inputs) > 8 else inputs_embeds
             output_attentions = inputs[9] if len(inputs) > 9 else output_attentions
             output_hidden_states = inputs[10] if len(inputs) > 10 else output_hidden_states
-            assert len(inputs) <= 11, "Too many inputs."
+            return_dict = inputs[11] if len(inputs) > 11 else return_dict
+            assert len(inputs) <= 12, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -168,12 +177,14 @@ class TFFlaubertMainLayer(TFXLMMainLayer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 11, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 12, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -260,8 +271,8 @@ class TFFlaubertMainLayer(TFXLMMainLayer):
         tensor = tensor * mask[..., tf.newaxis]
 
         # transformer layers
-        hidden_states = ()
+        hidden_states = () if output_hidden_states else None
-        attentions = ()
+        attentions = () if output_attentions else None
         for i in range(self.n_layers):
             # LayerDrop
             dropout_probability = random.uniform(0, 1)
@@ -321,12 +332,9 @@ class TFFlaubertMainLayer(TFXLMMainLayer):
         # move back sequence length to dimension 0
         # tensor = tensor.transpose(0, 1)
 
-        outputs = (tensor,)
+        if not return_dict:
-        if output_hidden_states:
+            return tuple(v for v in [tensor, hidden_states, attentions] if v is not None)
-            outputs = outputs + (hidden_states,)
+        return TFBaseModelOutput(last_hidden_state=tensor, hidden_states=hidden_states, attentions=attentions)
-        if output_attentions:
-            outputs = outputs + (attentions,)
-        return outputs  # outputs, (hidden_states), (attentions)
 
 
 @add_start_docstrings(

src/transformers/modeling_tf_gpt2.py

@@ -17,12 +17,21 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
 
 import numpy as np
 import tensorflow as tf
 
 from .configuration_gpt2 import GPT2Config
-from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
+from .file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_callable,
+    replace_return_docstrings,
+)
+from .modeling_tf_outputs import TFBaseModelOutputWithPast, TFCausalLMOutputWithPast
 from .modeling_tf_utils import (
     TFCausalLanguageModelingLoss,
     TFConv1D,
@@ -38,6 +47,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "GPT2Config"
 _TOKENIZER_FOR_DOC = "GPT2Tokenizer"
 
 TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -214,12 +224,11 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
         self.use_cache = config.use_cache
+        self.return_dict = config.use_return_dict
 
         self.num_hidden_layers = config.n_layer
         self.vocab_size = config.vocab_size
         self.n_embd = config.n_embd
-        self.output_hidden_states = self.output_hidden_states
-        self.output_attentions = self.output_attentions
 
         self.wte = TFSharedEmbeddings(
             config.vocab_size, config.hidden_size, initializer_range=config.initializer_range, name="wte"
@@ -259,6 +268,7 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -272,7 +282,8 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
             use_cache = inputs[7] if len(inputs) > 7 else use_cache
             output_attentions = inputs[8] if len(inputs) > 8 else output_attentions
             output_hidden_states = inputs[9] if len(inputs) > 9 else output_hidden_states
-            assert len(inputs) <= 10, "Too many inputs."
+            return_dict = inputs[10] if len(inputs) > 10 else return_dict
+            assert len(inputs) <= 11, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             past = inputs.get("past", past)
@@ -284,13 +295,15 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 10, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 11, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
         use_cache = use_cache if use_cache is not None else self.use_cache
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -355,9 +368,9 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
 
         output_shape = input_shape + [shape_list(hidden_states)[-1]]
 
-        presents = ()
+        presents = () if use_cache else None
-        all_attentions = []
+        all_attentions = () if output_attentions else None
-        all_hidden_states = ()
+        all_hidden_states = () if output_hidden_states else None
         for i, (block, layer_past) in enumerate(zip(self.h, past)):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
@@ -373,10 +386,11 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
             )
 
             hidden_states, present = outputs[:2]
-            presents = presents + (present,)
+            if use_cache:
+                presents = presents + (present,)
 
             if output_attentions:
-                all_attentions.append(outputs[2])
+                all_attentions = all_attentions + (outputs[2],)
 
         hidden_states = self.ln_f(hidden_states)
 
@@ -385,18 +399,20 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
         if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states,)
-
-        if use_cache:
-            outputs = outputs + (presents,)
-        if output_hidden_states:
-            outputs = outputs + (all_hidden_states,)
         if output_attentions:
             # let the number of heads free (-1) so we can extract attention even after head pruning
             attention_output_shape = input_shape[:-1] + [-1] + shape_list(all_attentions[0])[-2:]
             all_attentions = tuple(tf.reshape(t, attention_output_shape) for t in all_attentions)
-            outputs = outputs + (all_attentions,)
+
-        return outputs  # last hidden state, presents, (all hidden_states), (attentions)
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
 
 
 class TFGPT2PreTrainedModel(TFPreTrainedModel):
@@ -408,6 +424,42 @@ class TFGPT2PreTrainedModel(TFPreTrainedModel):
     base_model_prefix = "transformer"
 
 
+@dataclass
+class TFGPT2DoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        lm_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        mc_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            ``past_key_values`` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    lm_logits: tf.Tensor = None
+    mc_logits: tf.Tensor = None
+    past_key_values: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 GPT2_START_DOCSTRING = r"""
 
     .. note::
@@ -482,6 +534,11 @@ GPT2_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -495,29 +552,13 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
         self.transformer = TFGPT2MainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="gpt2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="gpt2",
+        output_type=TFBaseModelOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.GPT2Config`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the last layer of the model.
-        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-    """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -543,7 +584,12 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
         return {"inputs": inputs, "past": past, "use_cache": kwargs["use_cache"]}
 
     @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="gpt2")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="gpt2",
+        output_type=TFCausalLMOutputWithPast,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -556,6 +602,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -563,31 +610,12 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the cross entropy classification loss.
             Indices should be in ``[0, ..., config.vocab_size - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.GPT2Config`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[10] if len(inputs) > 10 else labels
+            labels = inputs[11] if len(inputs) > 11 else labels
-            if len(inputs) > 10:
+            if len(inputs) > 11:
-                inputs = inputs[:10]
+                inputs = inputs[:11]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -602,6 +630,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -609,15 +638,24 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel, TFCausalLanguageModelingLoss):
 
         logits = self.transformer.wte(hidden_states, mode="linear")
 
-        outputs = (logits,) + transformer_outputs[1:]
+        loss = None
         if labels is not None:
             # shift labels to the left and cut last logit token
             logits = logits[:, :-1]
             labels = labels[:, 1:]
             loss = self.compute_loss(labels, logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # lm_logits, presents, (all hidden_states), (attentions)
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -641,6 +679,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         return self.transformer.wte
 
     @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFGPT2DoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC)
     def call(
         self,
         inputs,
@@ -654,6 +693,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         r"""
@@ -662,26 +702,6 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             Selected in the range ``[0, input_ids.size(-1) - 1[``.
 
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.GPT2Config`) and inputs:
-        lm_prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        mc_prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
-            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
-        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as `input_ids` as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -717,8 +737,10 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             inputs_embeds = inputs[6] if len(inputs) > 6 else inputs_embeds
             mc_token_ids = inputs[7] if len(inputs) > 7 else mc_token_ids
             use_cache = inputs[8] if len(inputs) > 8 else use_cache
-            output_attentions = inputs[9] if len(inputs) > 8 else output_attentions
+            output_attentions = inputs[9] if len(inputs) > 9 else output_attentions
-            assert len(inputs) <= 10, "Too many inputs."
+            output_hidden_states = inputs[10] if len(inputs) > 10 else output_hidden_states
+            return_dict = inputs[11] if len(inputs) > 11 else return_dict
+            assert len(inputs) <= 12, "Too many inputs."
         elif isinstance(inputs, dict):
             input_ids = inputs.get("input_ids")
             past = inputs.get("past", past)
@@ -730,9 +752,12 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             mc_token_ids = inputs.get("mc_token_ids", mc_token_ids)
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
-            assert len(inputs) <= 10, "Too many inputs."
+            output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 12, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
 
         if input_ids is not None:
             input_shapes = shape_list(input_ids)
@@ -755,6 +780,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             use_cache,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         hidden_states = transformer_outputs[0]
@@ -762,6 +788,14 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         lm_logits = self.transformer.wte(hidden_states, mode="linear")
         mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids, training=training)
         mc_logits = tf.squeeze(mc_logits, axis=-1)
-        outputs = (lm_logits, mc_logits) + transformer_outputs[1:]
 
-        return outputs  # lm logits, mc logits, presents, (all hidden_states), (attentions)
+        if not return_dict:
+            return (lm_logits, mc_logits) + transformer_outputs[1:]
+
+        return TFGPT2DoubleHeadsModelOutput(
+            lm_logits=lm_logits,
+            mc_logits=mc_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

src/transformers/modeling_tf_mobilebert.py

@@ -17,17 +17,31 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import Optional, Tuple
 
 import tensorflow as tf
 
 from . import MobileBertConfig
 from .file_utils import (
     MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_callable,
+    replace_return_docstrings,
 )
 from .modeling_tf_bert import TFBertIntermediate, gelu, gelu_new, swish
+from .modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPooling,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFNextSentencePredictorOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMaskedLanguageModelingLoss,
     TFMultipleChoiceLoss,
@@ -44,6 +58,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "MobileBertConfig"
 _TOKENIZER_FOR_DOC = "MobileBertTokenizer"
 
 TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -541,9 +556,18 @@ class TFMobileBertEncoder(tf.keras.layers.Layer):
         self.output_hidden_states = config.output_hidden_states
         self.layer = [TFMobileBertLayer(config, name="layer_._{}".format(i)) for i in range(config.num_hidden_layers)]
 
-    def call(self, hidden_states, attention_mask, head_mask, output_attentions, output_hidden_states, training=False):
+    def call(
-        all_hidden_states = ()
+        self,
-        all_attentions = ()
+        hidden_states,
+        attention_mask,
+        head_mask,
+        output_attentions,
+        output_hidden_states,
+        return_dict,
+        training=False,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
         for i, layer_module in enumerate(self.layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
@@ -561,12 +585,11 @@ class TFMobileBertEncoder(tf.keras.layers.Layer):
         if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states,)
+        if not return_dict:
-        if output_hidden_states:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
-            outputs = outputs + (all_hidden_states,)
+        return TFBaseModelOutput(
-        if output_attentions:
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
-            outputs = outputs + (all_attentions,)
+        )
-        return outputs  # outputs, (hidden states), (attentions)
 
 
 class TFMobileBertPooler(tf.keras.layers.Layer):
@@ -660,6 +683,7 @@ class TFMobileBertMainLayer(tf.keras.layers.Layer):
         self.num_hidden_layers = config.num_hidden_layers
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
+        self.return_dict = config.use_return_dict
 
         self.embeddings = TFMobileBertEmbeddings(config, name="embeddings")
         self.encoder = TFMobileBertEncoder(config, name="encoder")
@@ -688,6 +712,7 @@ class TFMobileBertMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -699,7 +724,8 @@ class TFMobileBertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -709,12 +735,14 @@ class TFMobileBertMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -763,16 +791,22 @@ class TFMobileBertMainLayer(tf.keras.layers.Layer):
             head_mask,
             output_attentions,
             output_hidden_states,
+            return_dict,
             training=training,
         )
 
         sequence_output = encoder_outputs[0]
         pooled_output = self.pooler(sequence_output)
 
-        outputs = (sequence_output, pooled_output,) + encoder_outputs[
+        if not return_dict:
-            1:
+            return (sequence_output, pooled_output,) + encoder_outputs[1:]
-        ]  # add hidden_states and attentions if they are here
+
-        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
 
 
 class TFMobileBertPreTrainedModel(TFPreTrainedModel):
@@ -784,6 +818,37 @@ class TFMobileBertPreTrainedModel(TFPreTrainedModel):
     base_model_prefix = "mobilebert"
 
 
+@dataclass
+class TFMobileBertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFMobileBertForPreTrainingModel`.
+
+    Args:
+        prediction_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        seq_relationship_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False
+            continuation before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    prediction_logits: tf.Tensor = None
+    seq_relationship_logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 MOBILEBERT_START_DOCSTRING = r"""
     This model is a `tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ sub-class.
     Use it as a regular TF 2.0 Keras Model and
@@ -852,6 +917,13 @@ MOBILEBERT_INPUTS_DOCSTRING = r"""
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
+        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -865,32 +937,13 @@ class TFMobileBertModel(TFMobileBertPreTrainedModel):
         self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert")
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/mobilebert-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/mobilebert-uncased",
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`tf.Tensor` of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during the original Bert pretraining. This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.mobilebert(inputs, **kwargs)
         return outputs
 
@@ -911,25 +964,10 @@ class TFMobileBertForPreTraining(TFMobileBertPreTrainedModel):
         return self.mobilebert.embeddings
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @replace_return_docstrings(output_type=TFMobileBertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
     def call(self, inputs, **kwargs):
         r"""
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        seq_relationship_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, 2)`):
-            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -943,16 +981,23 @@ class TFMobileBertForPreTraining(TFMobileBertPreTrainedModel):
         >>> prediction_scores, seq_relationship_scores = outputs[:2]
 
         """
+        return_dict = kwargs.get("return_dict")
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
         outputs = self.mobilebert(inputs, **kwargs)
 
         sequence_output, pooled_output = outputs[:2]
         prediction_scores = self.predictions(sequence_output)
         seq_relationship_score = self.seq_relationship(pooled_output)
-        outputs = (prediction_scores, seq_relationship_score,) + outputs[
-            2:
-        ]  # add hidden states and attention if they are here
 
-        return outputs  # prediction_scores, seq_relationship_score, (hidden_states), (attentions)
+        if not return_dict:
+            return (prediction_scores, seq_relationship_score) + outputs[2:]
+
+        return TFMobileBertForPreTrainingOutput(
+            prediction_logits=prediction_scores,
+            seq_relationship_logits=seq_relationship_score,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings("""MobileBert Model with a `language modeling` head on top. """, MOBILEBERT_START_DOCSTRING)
@@ -967,7 +1012,12 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel
         return self.mobilebert.embeddings
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/mobilebert-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/mobilebert-uncased",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -978,6 +1028,7 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -986,27 +1037,12 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel
             Labels for computing the masked language modeling loss.
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1019,18 +1055,22 @@ class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModel
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
         sequence_output = outputs[0]
         prediction_scores = self.mlm(sequence_output, training=training)
 
-        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, prediction_scores)
-        if labels is not None:
-            loss = self.compute_loss(labels, prediction_scores)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), prediction_scores, (hidden_states), (attentions)
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMaskedLMOutput(
+            loss=loss, logits=prediction_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 class TFMobileBertOnlyNSPHead(tf.keras.layers.Layer):
@@ -1055,23 +1095,10 @@ class TFMobileBertForNextSentencePrediction(TFMobileBertPreTrainedModel):
         self.cls = TFMobileBertOnlyNSPHead(config, name="seq_relationship___cls")
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @replace_return_docstrings(output_type=TFNextSentencePredictorOutput, config_class=_CONFIG_FOR_DOC)
     def call(self, inputs, **kwargs):
         r"""
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        seq_relationship_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, 2)`)
-            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -1087,14 +1114,19 @@ class TFMobileBertForNextSentencePrediction(TFMobileBertPreTrainedModel):
 
         >>> logits = model(encoding['input_ids'], token_type_ids=encoding['token_type_ids'])[0]
         """
+        return_dict = kwargs.get("return_dict")
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
         outputs = self.mobilebert(inputs, **kwargs)
 
         pooled_output = outputs[1]
         seq_relationship_score = self.cls(pooled_output)
 
-        outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
+        if not return_dict:
+            return (seq_relationship_score,) + outputs[2:]
 
-        return outputs  # seq_relationship_score, (hidden_states), (attentions)
+        return TFNextSentencePredictorOutput(
+            logits=seq_relationship_score, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1114,7 +1146,12 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
         )
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/mobilebert-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/mobilebert-uncased",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1125,6 +1162,7 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1134,27 +1172,12 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
             Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
             If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
             If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1167,6 +1190,7 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1175,13 +1199,15 @@ class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSeque
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1200,7 +1226,12 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
         )
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/mobilebert-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/mobilebert-uncased",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1211,6 +1242,7 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -1224,30 +1256,13 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[8] if len(inputs) > 8 else start_positions
+            start_positions = inputs[9] if len(inputs) > 9 else start_positions
-            end_positions = inputs[9] if len(inputs) > 9 else end_positions
+            end_positions = inputs[10] if len(inputs) > 10 else end_positions
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -1261,6 +1276,7 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1271,15 +1287,23 @@ class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAn
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + outputs[2:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1307,7 +1331,12 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/mobilebert-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/mobilebert-uncased",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -1318,6 +1347,7 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1326,24 +1356,6 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -1354,8 +1366,9 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            labels = inputs[8] if len(inputs) > 8 else labels
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
-            assert len(inputs) <= 9, "Too many inputs."
+            labels = inputs[9] if len(inputs) > 9 else labels
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -1365,10 +1378,12 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 9, "Too many inputs."
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -1395,19 +1410,23 @@ class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoic
             flat_inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         pooled_output = outputs[1]
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
-            loss = self.compute_loss(labels, reshaped_logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1427,7 +1446,12 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
         )
 
     @add_start_docstrings_to_callable(MOBILEBERT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="google/mobilebert-uncased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="google/mobilebert-uncased",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1438,6 +1462,7 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1445,27 +1470,12 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.MobileBertConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.mobilebert.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1478,6 +1488,7 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1486,10 +1497,12 @@ class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenCla
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )

src/transformers/modeling_tf_openai.py

@@ -17,12 +17,21 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import Optional, Tuple
 
 import numpy as np
 import tensorflow as tf
 
 from .configuration_openai import OpenAIGPTConfig
-from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
+from .file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_callable,
+    replace_return_docstrings,
+)
+from .modeling_tf_outputs import TFBaseModelOutput, TFCausalLMOutput
 from .modeling_tf_utils import (
     TFCausalLanguageModelingLoss,
     TFConv1D,
@@ -38,6 +47,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "OpenAIGPTConfig"
 _TOKENIZER_FOR_DOC = "OpenAIGPTTokenizer"
 
 TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -208,6 +218,7 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
         super().__init__(*inputs, **kwargs)
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
+        self.return_dict = config.use_return_dict
         self.num_hidden_layers = config.n_layer
         self.vocab_size = config.vocab_size
         self.n_embd = config.n_embd
@@ -247,6 +258,7 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -258,7 +270,8 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
+            assert len(inputs) <= 9, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -268,12 +281,14 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 8, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -333,8 +348,8 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
 
         output_shape = input_shape + [shape_list(hidden_states)[-1]]
 
-        all_attentions = []
+        all_attentions = () if output_attentions else None
-        all_hidden_states = ()
+        all_hidden_states = () if output_hidden_states else None
         for i, block in enumerate(self.h):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
@@ -342,22 +357,24 @@ class TFOpenAIGPTMainLayer(tf.keras.layers.Layer):
             outputs = block(hidden_states, attention_mask, head_mask[i], output_attentions, training=training)
             hidden_states = outputs[0]
             if output_attentions:
-                all_attentions.append(outputs[1])
+                all_attentions = all_attentions + (outputs[1],)
 
         hidden_states = tf.reshape(hidden_states, output_shape)
         # Add last hidden state
         if output_hidden_states:
             all_hidden_states = all_hidden_states + (hidden_states,)
 
-        outputs = (hidden_states,)
-        if output_hidden_states:
-            outputs = outputs + (all_hidden_states,)
         if output_attentions:
             # let the number of heads free (-1) so we can extract attention even after head pruning
             attention_output_shape = input_shape[:-1] + [-1] + shape_list(all_attentions[0])[-2:]
             all_attentions = tuple(tf.reshape(t, attention_output_shape) for t in all_attentions)
-            outputs = outputs + (all_attentions,)
+
-        return outputs  # last hidden state, (all hidden_states), (attentions)
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions,
+        )
 
 
 class TFOpenAIGPTPreTrainedModel(TFPreTrainedModel):
@@ -369,6 +386,35 @@ class TFOpenAIGPTPreTrainedModel(TFPreTrainedModel):
     base_model_prefix = "transformer"
 
 
+@dataclass
+class TFOpenAIGPTDoubleHeadsModelOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        lm_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        mc_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
+            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    lm_logits: tf.Tensor = None
+    mc_logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 OPENAI_GPT_START_DOCSTRING = r"""
 
     .. note::
@@ -436,6 +482,11 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -449,25 +500,13 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
         self.transformer = TFOpenAIGPTMainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(OPENAI_GPT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="openai-gpt")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="openai-gpt",
+        output_type=TFBaseModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.OpenAIGPTConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the last layer of the model.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -486,7 +525,12 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
         return self.transformer.tokens_embed
 
     @add_start_docstrings_to_callable(OPENAI_GPT_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="openai-gpt")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="openai-gpt",
+        output_type=TFCausalLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -497,6 +541,7 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -504,27 +549,12 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the cross entropy classification loss.
             Indices should be in ``[0, ..., config.vocab_size - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.OpenAIGPTConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -537,21 +567,30 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel, TFCausalLanguageModelin
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         hidden_states = transformer_outputs[0]
 
         logits = self.transformer.tokens_embed(hidden_states, mode="linear")
-        outputs = (logits,) + transformer_outputs[1:]
 
+        loss = None
         if labels is not None:
             # shift labels to the left and cut last logit token
             logits = logits[:, :-1]
             labels = labels[:, 1:]
             loss = self.compute_loss(labels, logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # lm_logits, (all hidden_states), (attentions)
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFCausalLMOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -575,6 +614,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
         return self.transformer.tokens_embed
 
     @add_start_docstrings_to_callable(OPENAI_GPT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFOpenAIGPTDoubleHeadsModelOutput, config_class=_CONFIG_FOR_DOC)
     def call(
         self,
         inputs,
@@ -586,6 +626,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
         mc_token_ids=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         r"""
@@ -594,27 +635,6 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
             Selected in the range ``[0, input_ids.size(-1) - 1]``.
 
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.OpenAIGPTConfig`) and inputs:
-        lm_prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        mc_prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
-            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
-        past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-
 
     Examples::
 
@@ -646,7 +666,9 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             mc_token_ids = inputs[6] if len(inputs) > 6 else mc_token_ids
             output_attentions = inputs[7] if len(inputs) > 7 else output_attentions
-            assert len(inputs) <= 8, "Too many inputs."
+            output_hidden_states = inputs[8] if len(inputs) > 8 else output_hidden_states
+            return_dict = inputs[9] if len(inputs) > 9 else return_dict
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -656,9 +678,12 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             mc_token_ids = inputs.get("mc_token_ids", mc_token_ids)
             output_attentions = inputs.get("output_attentions", output_attentions)
-            assert len(inputs) <= 8, "Too many inputs."
+            output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
 
         if input_ids is not None:
             input_shapes = shape_list(input_ids)
@@ -679,6 +704,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         hidden_states = transformer_outputs[0]
@@ -686,6 +712,13 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
         lm_logits = self.transformer.tokens_embed(hidden_states, mode="linear")
         mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids, training=training)
         mc_logits = tf.squeeze(mc_logits, axis=-1)
-        outputs = (lm_logits, mc_logits) + transformer_outputs[1:]
 
-        return outputs  # lm logits, mc logits, (all hidden_states), (attentions)
+        if not return_dict:
+            return (lm_logits, mc_logits) + transformer_outputs[1:]
+
+        return TFOpenAIGPTDoubleHeadsModelOutput(
+            lm_logits=lm_logits,
+            mc_logits=mc_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

src/transformers/modeling_tf_outputs.py

@@ -0,0 +1,555 @@
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import tensorflow as tf
+
+from .file_utils import ModelOutput
+
+
+@dataclass
+class TFBaseModelOutput(ModelOutput):
+    """
+    Base class for model's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(tf.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFBaseModelOutputWithPooling(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`tf.Tensor` of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during pretraining.
+
+            This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: tf.Tensor = None
+    pooler_output: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFBaseModelOutputWithPast(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+
+            If `past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size, 1, hidden_size)` is output.
+        past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            ``past_key_values`` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: tf.Tensor = None
+    past_key_values: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFSeq2SeqModelOutput(ModelOutput):
+    """
+    Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential
+    decoding.
+
+    Args:
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the decoder of the model.
+
+            If ``decoder_past_key_values`` is used only the last hidden-state of the sequences of shape :obj:`(batch_size, 1, hidden_size)` is output.
+        decoder_past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
+            used (see ``decoder_past_key_values`` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        encoder_last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    last_hidden_state: tf.Tensor = None
+    decoder_past_key_values: Optional[List[tf.Tensor]] = None
+    decoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    decoder_attentions: Optional[Tuple[tf.Tensor]] = None
+    encoder_last_hidden_state: Optional[tf.Tensor] = None
+    encoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    encoder_attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFCausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFCausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            ``past_key_values`` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    past_key_values: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFMaskedLMOutput(ModelOutput):
+    """
+    Base class for masked language models outputs.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Masked languaged modeling (MLM) loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFSeq2SeqLMOutput(ModelOutput):
+    """
+    Base class for sequence-to-sequence language models outputs.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Languaged modeling loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        decoder_past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
+            used (see ``decoder_past_key_values`` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        encoder_last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    decoder_past_key_values: Optional[List[tf.Tensor]] = None
+    decoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    decoder_attentions: Optional[Tuple[tf.Tensor]] = None
+    encoder_last_hidden_state: Optional[tf.Tensor] = None
+    encoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    encoder_attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFNextSentencePredictorOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFSequenceClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of sentence classification models.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFSeq2SeqSequenceClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of sequence-to-sequence sentence classification models.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        decoder_past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
+            used (see ``decoder_past_key_values`` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        encoder_last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    decoder_past_key_values: Optional[List[tf.Tensor]] = None
+    decoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    decoder_attentions: Optional[Tuple[tf.Tensor]] = None
+    encoder_last_hidden_state: Optional[tf.Tensor] = None
+    encoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    encoder_attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFMultipleChoiceModelOutput(ModelOutput):
+    """
+    Base class for outputs of multiple choice models.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
+            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFTokenClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of token classification models.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFQuestionAnsweringModelOutput(ModelOutput):
+    """
+    Base class for outputs of question answering models.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-end scores (before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    start_logits: tf.Tensor = None
+    end_logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFSeq2SeqQuestionAnsweringModelOutput(ModelOutput):
+    """
+    Base class for outputs of sequence-to-sequence question answering models.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-end scores (before SoftMax).
+        decoder_past_key_values (:obj:`List[tf.Tensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            List of :obj:`tf.Tensor` of length :obj:`config.n_layers`,  with each tensor of shape
+            :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`).
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be
+            used (see ``decoder_past_key_values`` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        encoder_last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    start_logits: tf.Tensor = None
+    end_logits: tf.Tensor = None
+    decoder_past_key_values: Optional[List[tf.Tensor]] = None
+    decoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    decoder_attentions: Optional[Tuple[tf.Tensor]] = None
+    encoder_last_hidden_state: Optional[tf.Tensor] = None
+    encoder_hidden_states: Optional[Tuple[tf.Tensor]] = None
+    encoder_attentions: Optional[Tuple[tf.Tensor]] = None

src/transformers/modeling_tf_roberta.py

@@ -28,6 +28,14 @@ from .file_utils import (
     add_start_docstrings_to_callable,
 )
 from .modeling_tf_bert import TFBertEmbeddings, TFBertMainLayer, gelu
+from .modeling_tf_outputs import (
+    TFBaseModelOutputWithPooling,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMaskedLanguageModelingLoss,
     TFMultipleChoiceLoss,
@@ -44,6 +52,7 @@ from .tokenization_utils_base import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "RobertaConfig"
 _TOKENIZER_FOR_DOC = "RobertaTokenizer"
 
 TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -190,6 +199,11 @@ ROBERTA_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -203,32 +217,13 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
         self.roberta = TFRobertaMainLayer(config, name="roberta")
 
     @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="roberta-base")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="roberta-base",
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`tf.Tensor` of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during Bert pretraining. This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.roberta(inputs, **kwargs)
         return outputs
 
@@ -276,7 +271,12 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
         return self.lm_head.decoder
 
     @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="roberta-base")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="roberta-base",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -287,6 +287,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -296,27 +297,12 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.roberta.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -329,6 +315,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -337,13 +324,15 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
         sequence_output = outputs[0]
         prediction_scores = self.lm_head(sequence_output)
 
-        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, prediction_scores)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, prediction_scores)
+            output = (prediction_scores,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), prediction_scores, (hidden_states), (attentions)
+        return TFMaskedLMOutput(
+            loss=loss, logits=prediction_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 class TFRobertaClassificationHead(tf.keras.layers.Layer):
@@ -385,7 +374,12 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceCla
         self.classifier = TFRobertaClassificationHead(config, name="classifier")
 
     @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="roberta-base")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="roberta-base",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -396,30 +390,22 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceCla
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
         r"""
-    Return:
+        labels (:obj:`tf.Tensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
+            Labels for computing the sequence classification/regression loss.
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.roberta.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -432,19 +418,22 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceCla
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
         sequence_output = outputs[0]
         logits = self.classifier(sequence_output, training=training)
 
-        outputs = (logits,) + outputs[2:]
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -472,7 +461,12 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="roberta-base")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="roberta-base",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -483,6 +477,7 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -491,24 +486,6 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -519,8 +496,9 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            labels = inputs[8] if len(inputs) > 8 else labels
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
-            assert len(inputs) <= 9, "Too many inputs."
+            labels = inputs[9] if len(inputs) > 9 else labels
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -530,10 +508,12 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_attentions)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 9, "Too many inputs."
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.roberta.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -555,19 +535,23 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss)
             inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         pooled_output = outputs[1]
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
-            loss = self.compute_loss(labels, reshaped_logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -587,7 +571,12 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
         )
 
     @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="roberta-base")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="roberta-base",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -598,6 +587,7 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -605,27 +595,12 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.roberta.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -638,6 +613,7 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -646,13 +622,15 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -670,7 +648,12 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
         )
 
     @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="roberta-base")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="roberta-base",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -681,6 +664,7 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -694,30 +678,13 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.roberta.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[8] if len(inputs) > 8 else start_positions
+            start_positions = inputs[9] if len(inputs) > 9 else start_positions
-            end_positions = inputs[9] if len(inputs) > 9 else end_positions
+            end_positions = inputs[10] if len(inputs) > 10 else end_positions
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -731,6 +698,7 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -741,12 +709,20 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + outputs[2:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/modeling_tf_t5.py

@@ -25,7 +25,14 @@ import warnings
 import tensorflow as tf
 
 from .configuration_t5 import T5Config
-from .file_utils import DUMMY_INPUTS, DUMMY_MASK, add_start_docstrings, add_start_docstrings_to_callable
+from .file_utils import (
+    DUMMY_INPUTS,
+    DUMMY_MASK,
+    add_start_docstrings,
+    add_start_docstrings_to_callable,
+    replace_return_docstrings,
+)
+from .modeling_tf_outputs import TFSeq2SeqLMOutput, TFSeq2SeqModelOutput
 from .modeling_tf_utils import (
     TFCausalLanguageModelingLoss,
     TFPreTrainedModel,
@@ -39,6 +46,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "T5Config"
 _TOKENIZER_FOR_DOC = "T5Tokenizer"
 
 TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -575,8 +583,8 @@ class TFT5MainLayer(tf.keras.layers.Layer):
             head_mask = inputs[5] if len(inputs) > 5 else head_mask
             past_key_value_states = inputs[6] if len(inputs) > 6 else past_key_value_states
             use_cache = inputs[7] if len(inputs) > 7 else use_cache
-            output_attentions = inputs[8] if len(inputs) > 7 else output_attentions
+            output_attentions = inputs[8] if len(inputs) > 8 else output_attentions
-            output_hidden_states = inputs[9] if len(inputs) > 8 else output_hidden_states
+            output_hidden_states = inputs[9] if len(inputs) > 9 else output_hidden_states
             assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
@@ -934,6 +942,7 @@ class TFT5Model(TFT5PreTrainedModel):
         return self.decoder
 
     @add_start_docstrings_to_callable(T5_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
     def call(
         self,
         inputs,
@@ -948,29 +957,11 @@ class TFT5Model(TFT5PreTrainedModel):
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         r"""
     Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.T5Config`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-            If `decoder_past_key_value_states` is used only the last hidden-state of the sequences of shape :obj:`(batch_size, 1, hidden_size)` is output.
-        decoder_past_key_value_states (:obj:`tuple(tuple(tf.Tensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`, `optional`, returned when ``use_cache=True``):
-            Contains pre-computed key and value hidden-states of the attention blocks.
-            Can be used to speed up sequential decoding (see `decoder_past_key_value_states` input).
-            Note that when using `decoder_past_key_value_states`, the model only outputs the last `hidden-state` of the sequence of shape :obj:`(batch_size, 1, config.vocab_size)`.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -996,7 +987,8 @@ class TFT5Model(TFT5PreTrainedModel):
             use_cache = inputs[9] if len(inputs) > 9 else use_cache
             output_attentions = inputs[10] if len(inputs) > 10 else output_attentions
             output_hidden_states = inputs[11] if len(inputs) > 11 else output_hidden_states
-            assert len(inputs) <= 12, "Too many inputs."
+            return_dict = inputs[12] if len(inputs) > 12 else return_dict
+            assert len(inputs) <= 13, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             if "inputs" in inputs:
                 warnings.warn("Using `inputs` as a keyword argument is deprecated. Please use `input_ids` instead.")
@@ -1013,11 +1005,13 @@ class TFT5Model(TFT5PreTrainedModel):
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 12, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 13, "Too many inputs."
         else:
             input_ids = inputs
 
         use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
 
         # Encode if needed (training, first prediction pass)
         if encoder_outputs is None:
@@ -1063,12 +1057,40 @@ class TFT5Model(TFT5PreTrainedModel):
             ],
             training=training,
         )
+        past = (
+            (encoder_outputs, decoder_outputs[1]) if cast_bool_to_primitive(use_cache, self.config.use_cache) else None
+        )
+        if not return_dict:
+            if past is not None:
+                decoder_outputs = decoder_outputs[:1] + (past,) + decoder_outputs[2:]
+            return decoder_outputs + encoder_outputs
 
-        if cast_bool_to_primitive(use_cache, self.config.use_cache) is True:
+        # If put before, this breaks the tf compilation.
-            past = ((encoder_outputs, decoder_outputs[1]),)
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-            decoder_outputs = decoder_outputs[:1] + past + decoder_outputs[2:]
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
 
-        return decoder_outputs + encoder_outputs
+        # This is long and annoying but if we introduce return_dict at the TFT5MainLayer level (like in PyTorch)
+        # TF refuses to compile anymore.
+        if not cast_bool_to_primitive(use_cache, self.config.use_cache):
+            decoder_outputs = decoder_outputs[:1] + (None,) + decoder_outputs[1:]
+        if not cast_bool_to_primitive(output_hidden_states, self.config.output_hidden_states):
+            encoder_outputs = encoder_outputs[:1] + (None,) + encoder_outputs[1:]
+            decoder_outputs = decoder_outputs[:2] + (None,) + decoder_outputs[2:]
+        if not cast_bool_to_primitive(output_attentions, self.config.output_attentions):
+            encoder_outputs = encoder_outputs + (None,)
+            decoder_outputs = decoder_outputs + (None,)
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs[0],
+            decoder_past_key_values=past,
+            decoder_hidden_states=decoder_outputs[2],
+            decoder_attentions=decoder_outputs[3],
+            encoder_last_hidden_state=encoder_outputs[0],
+            encoder_hidden_states=encoder_outputs[1],
+            encoder_attentions=encoder_outputs[2],
+        )
 
 
 @add_start_docstrings("""T5 Model with a `language modeling` head on top. """, T5_START_DOCSTRING)
@@ -1115,6 +1137,7 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
         return self.decoder
 
     @add_start_docstrings_to_callable(T5_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
     def call(
         self,
         inputs,
@@ -1129,6 +1152,7 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
         use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1138,24 +1162,6 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
             Indices should be in ``[0, ..., config.vocab_size - 1]``.
 
     Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.T5Config`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        decoder_past_key_value_states (:obj:`tuple(tuple(tf.Tensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`, `optional`, returned when ``use_cache=True``):
-            Contains pre-computed key and value hidden-states of the attention blocks.
-            Can be used to speed up sequential decoding (see `decoder_past_key_value_states` input).
-            Note that when using `decoder_past_key_value_states`, the model only outputs the last `prediction_score` of the sequence of shape :obj:`(batch_size, 1, config.vocab_size)`.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -1186,8 +1192,9 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
             use_cache = inputs[9] if len(inputs) > 9 else use_cache
             output_attentions = inputs[10] if len(inputs) > 10 else output_attentions
             output_hidden_states = inputs[11] if len(inputs) > 11 else output_hidden_states
-            labels = inputs[12] if len(inputs) > 12 else labels
+            return_dict = inputs[12] if len(inputs) > 12 else return_dict
-            assert len(inputs) <= 13, "Too many inputs."
+            labels = inputs[13] if len(inputs) > 13 else labels
+            assert len(inputs) <= 14, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             if "inputs" in inputs:
                 warnings.warn("Using `inputs` as a keyword argument is deprecated. Please use `input_ids` instead.")
@@ -1204,12 +1211,14 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 13, "Too many inputs."
+            assert len(inputs) <= 14, "Too many inputs."
         else:
             input_ids = inputs
 
         use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
 
         # Encode if needed (training, first prediction pass)
         if encoder_outputs is None:
@@ -1261,22 +1270,48 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
             training=training,
         )
 
-        # insert decoder past at right place
-        # to speed up decoding
-        if cast_bool_to_primitive(use_cache, self.config.use_cache) is True:
-            past = ((encoder_outputs, decoder_outputs[1]),)
-            decoder_outputs = decoder_outputs[:1] + past + decoder_outputs[2:]
-
         sequence_output = decoder_outputs[0] * (self.model_dim ** -0.5)
         embed_tokens = self.get_output_embeddings()
         logits = embed_tokens(sequence_output, mode="linear")
-        decoder_outputs = (logits,) + decoder_outputs[1:]
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, logits)
-            loss = self.compute_loss(labels, logits)
-            decoder_outputs = (loss,) + decoder_outputs
 
-        return decoder_outputs + encoder_outputs
+        past = (
+            (encoder_outputs, decoder_outputs[1]) if cast_bool_to_primitive(use_cache, self.config.use_cache) else None
+        )
+        if not return_dict:
+            if past is not None:
+                decoder_outputs = decoder_outputs[:1] + (past,) + decoder_outputs[2:]
+            output = (logits,) + decoder_outputs[1:] + encoder_outputs
+            return ((loss,) + output) if loss is not None else output
+
+        # Putting this before breaks tf compilation.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        # This is long and annoying but if we introduce return_dict at the TFT5MainLayer level (like in PyTorch)
+        # TF refuses to compile anymore.
+        if not cast_bool_to_primitive(use_cache, self.config.use_cache):
+            decoder_outputs = decoder_outputs[:1] + (None,) + decoder_outputs[1:]
+        if not cast_bool_to_primitive(output_hidden_states, self.config.output_hidden_states):
+            encoder_outputs = encoder_outputs[:1] + (None,) + encoder_outputs[1:]
+            decoder_outputs = decoder_outputs[:2] + (None,) + decoder_outputs[2:]
+        if not cast_bool_to_primitive(output_attentions, self.config.output_attentions):
+            encoder_outputs = encoder_outputs + (None,)
+            decoder_outputs = decoder_outputs + (None,)
+
+        return TFSeq2SeqLMOutput(
+            loss=loss,
+            logits=logits,
+            decoder_past_key_values=past,
+            decoder_hidden_states=decoder_outputs[2],
+            decoder_attentions=decoder_outputs[3],
+            encoder_last_hidden_state=encoder_outputs[0],
+            encoder_hidden_states=encoder_outputs[1],
+            encoder_attentions=encoder_outputs[2],
+        )
 
     def prepare_inputs_for_generation(self, inputs, past, attention_mask, use_cache, **kwargs):
         assert past is not None, "past has to be defined for encoder_outputs"

src/transformers/modeling_tf_transfo_xl.py

@@ -18,11 +18,13 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
 
 import tensorflow as tf
 
 from .configuration_transfo_xl import TransfoXLConfig
-from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
+from .file_utils import ModelOutput, add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
 from .modeling_tf_transfo_xl_utilities import TFAdaptiveSoftmaxMask
 from .modeling_tf_utils import TFPreTrainedModel, get_initializer, keras_serializable, shape_list
 from .tokenization_utils import BatchEncoding
@@ -30,6 +32,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "TransfoXLConfig"
 _TOKENIZER_FOR_DOC = "TransfoXLTokenizer"
 
 TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -388,6 +391,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         super().__init__(**kwargs)
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
+        self.return_dict = config.use_return_dict
 
         self.n_token = config.vocab_size
 
@@ -525,6 +529,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -533,8 +538,9 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
             head_mask = inputs[2] if len(inputs) > 2 else head_mask
             inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
             output_attentions = inputs[4] if len(inputs) > 4 else output_attentions
-            output_hidden_states = inputs[5] if len(inputs) > 4 else output_hidden_states
+            output_hidden_states = inputs[5] if len(inputs) > 5 else output_hidden_states
-            assert len(inputs) <= 6, "Too many inputs."
+            return_dict = inputs[6] if len(inputs) > 6 else return_dict
+            assert len(inputs) <= 7, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             mems = inputs.get("mems", mems)
@@ -542,12 +548,14 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 6, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 7, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         # the original code for Transformer-XL used shapes [len, bsz] but we want a unified interface in the library
         # so we transpose here from shape [bsz, len] to shape [len, bsz]
@@ -606,7 +614,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         #         word_emb.new_ones((qlen, klen), dtype=torch.uint8), diagonal=1+mlen)[:,:,None]
 
         hids = []
-        attentions = []
+        attentions = [] if output_attentions else None
         if self.attn_type == 0:  # default
             pos_seq = tf.range(klen - 1, -1, -1.0)
             if self.clamp_len > 0:
@@ -633,17 +641,24 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         new_mems = self._update_mems(hids, mems, mlen, qlen)
 
         # We transpose back here to shape [bsz, len, hidden_dim]
-        outputs = [tf.transpose(core_out, perm=(1, 0, 2)), new_mems]
+        core_out = tf.transpose(core_out, perm=(1, 0, 2))
+
         if output_hidden_states:
             # Add last layer and transpose to library standard shape [bsz, len, hidden_dim]
             hids.append(core_out)
-            hids = list(tf.transpose(t, perm=(1, 0, 2)) for t in hids)
+            hids = tuple(tf.transpose(t, perm=(1, 0, 2)) for t in hids)
-            outputs.append(hids)
+        else:
+            hids = None
         if output_attentions:
             # Transpose to library standard shape [bsz, n_heads, query_seq_len, key_seq_len]
-            attentions = list(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
+            attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
-            outputs.append(attentions)
+
-        return outputs  # last hidden state, new_mems, (all hidden states), (all attentions)
+        if not return_dict:
+            return tuple(v for v in [core_out, new_mems, hids, attentions] if v is not None)
+
+        return TFTransfoXLModelOutput(
+            last_hidden_state=core_out, mems=new_mems, hidden_states=hids, attentions=attentions,
+        )
 
 
 class TFTransfoXLPreTrainedModel(TFPreTrainedModel):
@@ -655,6 +670,70 @@ class TFTransfoXLPreTrainedModel(TFPreTrainedModel):
     base_model_prefix = "transformer"
 
 
+@dataclass
+class TFTransfoXLModelOutput(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states (key and values in the attention blocks).
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: tf.Tensor = None
+    mems: List[tf.Tensor] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFTransfoXLLMHeadModelOutput(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        losses (:obj:`tf.Tensor` of shape `(batch_size, sequence_length-1)`, `optional`, returned when ``labels`` is provided)
+            Language modeling losses (not reduced).
+        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token after SoftMax).
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states (key and values in the attention blocks).
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    prediction_scores: tf.Tensor = None
+    mems: List[tf.Tensor] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 TRANSFO_XL_START_DOCSTRING = r"""
 
     .. note::
@@ -706,6 +785,11 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             than the model's internal embedding lookup matrix.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -719,29 +803,13 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
         self.transformer = TFTransfoXLMainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(TRANSFO_XL_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="transfo-xl-wt103")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="transfo-xl-wt103",
+        output_type=TFTransfoXLModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.TransfoXLConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the last layer of the model.
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -797,57 +865,47 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
         return self.transformer.init_mems(bsz)
 
     @add_start_docstrings_to_callable(TRANSFO_XL_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="transfo-xl-wt103")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="transfo-xl-wt103",
+        output_type=TFTransfoXLLMHeadModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
         mems=None,
         head_mask=None,
         inputs_embeds=None,
-        labels=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
+        labels=None,
         training=False,
     ):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.TransfoXLConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             mems = inputs[1] if len(inputs) > 1 else mems
             head_mask = inputs[2] if len(inputs) > 2 else head_mask
             inputs_embeds = inputs[3] if len(inputs) > 3 else inputs_embeds
-            labels = inputs[4] if len(inputs) > 4 else labels
+            output_attentions = inputs[4] if len(inputs) > 4 else output_attentions
-            output_attentions = inputs[5] if len(inputs) > 5 else output_attentions
+            output_hidden_states = inputs[5] if len(inputs) > 5 else output_hidden_states
-            assert len(inputs) <= 6, "Too many inputs."
+            return_dict = inputs[6] if len(inputs) > 6 else return_dict
+            labels = inputs[7] if len(inputs) > 7 else labels
+            assert len(inputs) <= 8, "Too many inputs."
         elif isinstance(inputs, (BatchEncoding, dict)):
             input_ids = inputs.get("input_ids")
             mems = inputs.get("mems", mems)
             head_mask = inputs.get("head_mask", head_mask)
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
-            labels = inputs.get("labels", labels)
             output_attentions = inputs.get("output_attentions", output_attentions)
-            assert len(inputs) <= 6, "Too many inputs."
+            output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
+            labels = inputs.get("labels", labels)
+            assert len(inputs) <= 8, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
 
         if input_ids is not None:
             bsz, tgt_len = shape_list(input_ids)[:2]
@@ -855,17 +913,30 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
             bsz, tgt_len = shape_list(inputs_embeds)[:2]
 
         transformer_outputs = self.transformer(
-            input_ids, mems, head_mask, inputs_embeds, output_attentions, output_hidden_states, training=training
+            input_ids,
+            mems,
+            head_mask,
+            inputs_embeds,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            training=training,
         )
 
         last_hidden = transformer_outputs[0]
         pred_hid = last_hidden[:, -tgt_len:]
-        outputs = transformer_outputs[1:]
 
         softmax_output = self.crit(pred_hid, labels, training=training)
-        outputs = [softmax_output] + outputs
 
-        return outputs  # logits, new_mems, (all hidden states), (all attentions)
+        if not return_dict:
+            return (softmax_output,) + transformer_outputs[1:]
+
+        return TFTransfoXLLMHeadModelOutput(
+            prediction_scores=softmax_output,
+            mems=transformer_outputs.mems,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
     def prepare_inputs_for_generation(self, inputs, past, **model_kwargs):
         inputs = {"inputs": inputs}

src/transformers/modeling_tf_xlm.py

@@ -20,6 +20,8 @@ import itertools
 import logging
 import math
 import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple
 
 import numpy as np
 import tensorflow as tf
@@ -27,10 +29,18 @@ import tensorflow as tf
 from .configuration_xlm import XLMConfig
 from .file_utils import (
     MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_callable,
 )
+from .modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMultipleChoiceLoss,
     TFPreTrainedModel,
@@ -48,6 +58,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "XLMConfig"
 _TOKENIZER_FOR_DOC = "XLMTokenizer"
 
 TF_XLM_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -224,6 +235,7 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         super().__init__(**kwargs)
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
+        self.return_dict = config.use_return_dict
 
         # encoder / decoder, output layer
         self.is_encoder = config.is_encoder
@@ -340,6 +352,7 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):  # removed: src_enc=None, src_len=None
         if isinstance(inputs, (tuple, list)):
@@ -354,7 +367,8 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs[8] if len(inputs) > 8 else inputs_embeds
             output_attentions = inputs[9] if len(inputs) > 9 else output_attentions
             output_hidden_states = inputs[10] if len(inputs) > 10 else output_hidden_states
-            assert len(inputs) <= 11, "Too many inputs."
+            return_dict = inputs[11] if len(inputs) > 11 else return_dict
+            assert len(inputs) <= 12, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -367,12 +381,14 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 11, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 12, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -454,8 +470,8 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         tensor = tensor * mask[..., tf.newaxis]
 
         # transformer layers
-        hidden_states = ()
+        hidden_states = () if output_hidden_states else None
-        attentions = ()
+        attentions = () if output_attentions else None
         for i in range(self.n_layers):
             if output_hidden_states:
                 hidden_states = hidden_states + (tensor,)
@@ -494,12 +510,9 @@ class TFXLMMainLayer(tf.keras.layers.Layer):
         # move back sequence length to dimension 0
         # tensor = tensor.transpose(0, 1)
 
-        outputs = (tensor,)
+        if not return_dict:
-        if output_hidden_states:
+            return tuple(v for v in [tensor, hidden_states, attentions] if v is not None)
-            outputs = outputs + (hidden_states,)
+        return TFBaseModelOutput(last_hidden_state=tensor, hidden_states=hidden_states, attentions=attentions)
-        if output_attentions:
-            outputs = outputs + (attentions,)
-        return outputs  # outputs, (hidden_states), (attentions)
 
 
 class TFXLMPreTrainedModel(TFPreTrainedModel):
@@ -522,6 +535,33 @@ class TFXLMPreTrainedModel(TFPreTrainedModel):
         return {"input_ids": inputs_list, "attention_mask": attns_list, "langs": langs_list}
 
 
+# Remove when XLMWithLMHead computes loss like other LM models
+@dataclass
+class TFXLMWithLMHeadModelOutput(ModelOutput):
+    """
+    Base class for :class:`~transformers.TFXLMWithLMHeadModel` outputs.
+
+    Args:
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    logits: tf.Tensor = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 XLM_START_DOCSTRING = r"""
 
     .. note::
@@ -603,6 +643,11 @@ XLM_INPUTS_DOCSTRING = r"""
             than the model's internal embedding lookup matrix.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -616,25 +661,13 @@ class TFXLMModel(TFXLMPreTrainedModel):
         self.transformer = TFXLMMainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlm-mlm-en-2048")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlm-mlm-en-2048",
+        output_type=TFBaseModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLMConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`tf.Tensor` or :obj:`Numpy array` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
-            Tuple of :obj:`tf.Tensor` or :obj:`Numpy array` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -701,32 +734,26 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
         return {"inputs": inputs, "langs": langs}
 
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlm-mlm-en-2048")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlm-mlm-en-2048",
+        output_type=TFXLMWithLMHeadModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
+        return_dict = kwargs.get("return_dict")
-    Return:
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLMConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         transformer_outputs = self.transformer(inputs, **kwargs)
 
         output = transformer_outputs[0]
         outputs = self.pred_layer(output)
-        outputs = (outputs,) + transformer_outputs[1:]  # Keep new_mems and attention/hidden states if they are here
 
-        return outputs
+        if not return_dict:
+            return (outputs,) + transformer_outputs[1:]
+
+        return TFXLMWithLMHeadModelOutput(
+            logits=outputs, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions
+        )
 
 
 @add_start_docstrings(
@@ -743,7 +770,12 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
         self.sequence_summary = TFSequenceSummary(config, initializer_range=config.init_std, name="sequence_summary")
 
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlm-mlm-en-2048")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlm-mlm-en-2048",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -757,6 +789,7 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -766,27 +799,12 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
             Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLMConfig`) and inputs:
-        logits (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[11] if len(inputs) > 11 else labels
+            labels = inputs[12] if len(inputs) > 12 else labels
-            if len(inputs) > 11:
+            if len(inputs) > 12:
-                inputs = inputs[:11]
+                inputs = inputs[:12]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -802,19 +820,25 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         output = transformer_outputs[0]
 
         logits = self.sequence_summary(output)
 
-        outputs = (logits,) + transformer_outputs[1:]  # Keep new_mems and attention/hidden states if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + transformer_outputs[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -845,7 +869,12 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
         }
 
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlm-mlm-en-2048")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlm-mlm-en-2048",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -859,6 +888,7 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -867,24 +897,6 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -898,8 +910,9 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs[8] if len(inputs) > 8 else inputs_embeds
             output_attentions = inputs[9] if len(inputs) > 9 else output_attentions
             output_hidden_states = inputs[10] if len(inputs) > 10 else output_hidden_states
-            labels = inputs[11] if len(inputs) > 11 else labels
+            return_dict = inputs[11] if len(inputs) > 11 else return_dict
-            assert len(inputs) <= 12, "Too many inputs."
+            labels = inputs[12] if len(inputs) > 12 else labels
+            assert len(inputs) <= 13, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -912,10 +925,12 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 12, "Too many inputs."
+            assert len(inputs) <= 13, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -955,19 +970,26 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, TFMultipleChoiceLoss):
             flat_inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         output = transformer_outputs[0]
         logits = self.sequence_summary(output)
         logits = self.logits_proj(logits)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + transformer_outputs[1:]  # add hidden states and attention if they are here
 
-        if labels is not None:
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
-            loss = self.compute_loss(labels, reshaped_logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (reshaped_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -987,7 +1009,12 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
         )
 
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlm-mlm-en-2048")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlm-mlm-en-2048",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1001,6 +1028,7 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1008,27 +1036,12 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[11] if len(inputs) > 11 else labels
+            labels = inputs[12] if len(inputs) > 12 else labels
-            if len(inputs) > 11:
+            if len(inputs) > 12:
-                inputs = inputs[:11]
+                inputs = inputs[:12]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1044,6 +1057,7 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1052,13 +1066,18 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + transformer_outputs[1:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + transformer_outputs[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1075,7 +1094,12 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
         )
 
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlm-mlm-en-2048")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlm-mlm-en-2048",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1089,6 +1113,7 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -1102,30 +1127,13 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLMConfig`) and inputs:
-        start_scores (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[11] if len(inputs) > 11 else start_positions
+            start_positions = inputs[12] if len(inputs) > 12 else start_positions
-            end_positions = inputs[12] if len(inputs) > 12 else end_positions
+            end_positions = inputs[13] if len(inputs) > 13 else end_positions
-            if len(inputs) > 11:
+            if len(inputs) > 12:
-                inputs = inputs[:11]
+                inputs = inputs[:12]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -1142,6 +1150,7 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1152,14 +1161,20 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, TFQuestionAnsweringL
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + transformer_outputs[
+        loss = None
-            1:
-        ]  # Keep mems, hidden states, attentions if there are in it
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

src/transformers/modeling_tf_xlm_roberta.py

@@ -62,8 +62,6 @@ XLM_ROBERTA_START_DOCSTRING = r"""
         config (:class:`~transformers.XLMRobertaConfig`): Model configuration class with all the parameters of the
             model. Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
-        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
-            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 

src/transformers/modeling_tf_xlnet.py

@@ -18,6 +18,8 @@
 
 
 import logging
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
 
 import numpy as np
 import tensorflow as tf
@@ -25,9 +27,11 @@ import tensorflow as tf
 from .configuration_xlnet import XLNetConfig
 from .file_utils import (
     MULTIPLE_CHOICE_DUMMY_INPUTS,
+    ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_callable,
+    replace_return_docstrings,
 )
 from .modeling_tf_utils import (
     TFCausalLanguageModelingLoss,
@@ -47,6 +51,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "XLNetConfig"
 _TOKENIZER_FOR_DOC = "XLNetTokenizer"
 
 TF_XLNET_PRETRAINED_MODEL_ARCHIVE_LIST = [
@@ -436,6 +441,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         super().__init__(**kwargs)
         self.output_hidden_states = config.output_hidden_states
         self.output_attentions = config.output_attentions
+        self.return_dict = config.return_dict
 
         self.mem_len = config.mem_len
         self.reuse_len = config.reuse_len
@@ -586,6 +592,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         use_cache=True,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         training=False,
     ):
         if isinstance(inputs, (tuple, list)):
@@ -601,7 +608,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             use_cache = inputs[9] if len(inputs) > 9 else use_cache
             output_attentions = inputs[10] if len(inputs) > 10 else output_attentions
             output_hidden_states = inputs[11] if len(inputs) > 11 else output_hidden_states
-            assert len(inputs) <= 12, "Too many inputs."
+            return_dict = inputs[12] if len(inputs) > 12 else return_dict
+            assert len(inputs) <= 13, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -615,12 +623,14 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
-            assert len(inputs) <= 12, "Too many inputs."
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 13, "Too many inputs."
         else:
             input_ids = inputs
 
         output_attentions = output_attentions if output_attentions is not None else self.output_attentions
         output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
 
         # the original code for XLNet uses shapes [len, bsz] with the batch dimension at the end
         # but we want a unified interface in the library with the batch size on the first dimension
@@ -743,8 +753,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         if mems is None:
             mems = [None] * len(self.layer)
 
-        attentions = []
+        attentions = [] if output_attentions else None
-        hidden_states = []
+        hidden_states = [] if output_hidden_states else None
         for i, layer_module in enumerate(self.layer):
             # cache new mems
             if self.mem_len is not None and self.mem_len > 0 and use_cache:
@@ -776,22 +786,24 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         output = self.dropout(output_g if output_g is not None else output_h, training=training)
 
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
-        outputs = (tf.transpose(output, perm=(1, 0, 2)),)
+        output = tf.transpose(output, perm=(1, 0, 2))
-
-        if self.mem_len is not None and self.mem_len > 0 and use_cache:
-            outputs = outputs + (new_mems,)
 
+        if not (self.mem_len is not None and self.mem_len > 0 and use_cache):
+            new_mems = None
         if output_hidden_states:
             if output_g is not None:
                 hidden_states = tuple(tf.transpose(h, perm=(1, 0, 2)) for hs in hidden_states for h in hs)
             else:
                 hidden_states = tuple(tf.transpose(hs, perm=(1, 0, 2)) for hs in hidden_states)
-            outputs = outputs + (hidden_states,)
         if output_attentions:
             attentions = tuple(tf.transpose(t, perm=(2, 3, 0, 1)) for t in attentions)
-            outputs = outputs + (attentions,)
 
-        return outputs  # outputs, (new_mems), (hidden_states), (attentions)
+        if not return_dict:
+            return tuple(v for v in [output, new_mems, hidden_states, attentions] if v is not None)
+
+        return TFXLNetModelOutput(
+            last_hidden_state=output, mems=new_mems, hidden_states=hidden_states, attentions=attentions
+        )
 
 
 class TFXLNetPreTrainedModel(TFPreTrainedModel):
@@ -803,6 +815,218 @@ class TFXLNetPreTrainedModel(TFPreTrainedModel):
     base_model_prefix = "transformer"
 
 
+@dataclass
+class TFXLNetModelOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFXLNetModel`.
+
+    Args:
+        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_predict, hidden_size)`):
+            Sequence of hidden-states at the last layer of the model.
+
+            ``num_predict`` corresponds to ``target_mapping.shape[1]``. If ``target_mapping`` is ``None``, then
+            ``num_predict`` corresponds to ``sequence_length``.
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states.
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: tf.Tensor = None
+    mems: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFXLNetLMHeadModelOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFXLNetLMHeadModel`.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape `(1,)`, `optional`, returned when ``labels`` is provided)
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_predict, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+
+            ``num_predict`` corresponds to ``target_mapping.shape[1]``. If ``target_mapping`` is ``None``, then
+            ``num_predict`` corresponds to ``sequence_length``.
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states.
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    mems: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFXLNetForSequenceClassificationOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFXLNetForSequenceClassification`.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states.
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    mems: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFXLNetForTokenClassificationOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFXLNetForTokenClassificationOutput`.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
+            Classification scores (before SoftMax).
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states.
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    mems: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFXLNetForMultipleChoiceOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFXLNetForMultipleChoice`.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification loss.
+        logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`):
+            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states.
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    logits: tf.Tensor = None
+    mems: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
+@dataclass
+class TFXLNetForQuestionAnsweringSimpleOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.TFXLNetForQuestionAnsweringSimple`.
+
+    Args:
+        loss (:obj:`tf.Tensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-end scores (before SoftMax).
+        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
+            Contains pre-computed hidden-states.
+            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
+        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`tf.Tensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[tf.Tensor] = None
+    start_logits: tf.Tensor = None
+    end_logits: tf.Tensor = None
+    mems: Optional[List[tf.Tensor]] = None
+    hidden_states: Optional[Tuple[tf.Tensor]] = None
+    attentions: Optional[Tuple[tf.Tensor]] = None
+
+
 XLNET_START_DOCSTRING = r"""
 
     .. note::
@@ -885,6 +1109,11 @@ XLNET_INPUTS_DOCSTRING = r"""
             If `use_cache` is True, `mems` are returned and can be used to speed up decoding (see `mems`). Defaults to `True`.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -898,29 +1127,13 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
         self.transformer = TFXLNetMainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlnet-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlnet-base-cased",
+        output_type=TFXLNetModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the last layer of the model.
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `mems` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`tf.Tensor` or :obj:`Numpy array` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
-            Tuple of :obj:`tf.Tensor` or :obj:`Numpy array` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -980,6 +1193,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
         return inputs
 
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFXLNetLMHeadModelOutput, config_class=_CONFIG_FOR_DOC)
     def call(
         self,
         inputs,
@@ -994,6 +1208,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
         use_cache=True,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1003,24 +1218,6 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
             Indices should be in ``[0, ..., config.vocab_size - 1]``.
 
     Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
-        prediction_scores (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
 
     Examples::
 
@@ -1045,10 +1242,11 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
         next_token_logits = outputs[0]  # Output has shape [target_mapping.size(0), target_mapping.size(1), config.vocab_size]
 
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[12] if len(inputs) > 12 else labels
+            labels = inputs[13] if len(inputs) > 13 else labels
-            if len(inputs) > 12:
+            if len(inputs) > 13:
-                inputs = inputs[:12]
+                inputs = inputs[:13]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1065,21 +1263,30 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel, TFCausalLanguageModelingLoss):
             use_cache=True,
             output_attentions=None,
             output_hidden_states=None,
+            return_dict=return_dict,
             training=training,
         )
         hidden_state = transformer_outputs[0]
         logits = self.lm_loss(hidden_state, training=training)
 
-        outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
+        loss = None
-
         if labels is not None:
             # shift labels to the left and cut last logit token
             logits = logits[:, :-1]
             labels = labels[:, 1:]
             loss = self.compute_loss(labels, logits)
-            outputs = (loss,) + outputs
 
-        return outputs  # return logits, (mems), (hidden states), (attentions)
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFXLNetLMHeadModelOutput(
+            loss=loss,
+            logits=logits,
+            mems=transformer_outputs.mems,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1101,7 +1308,12 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
         )
 
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlnet-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlnet-base-cased",
+        output_type=TFXLNetForSequenceClassificationOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1116,6 +1328,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
         use_cache=True,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1125,31 +1338,12 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
             Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
-        logits (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[12] if len(inputs) > 12 else labels
+            labels = inputs[13] if len(inputs) > 13 else labels
-            if len(inputs) > 12:
+            if len(inputs) > 13:
-                inputs = inputs[:12]
+                inputs = inputs[:13]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1166,19 +1360,26 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
         )
         output = transformer_outputs[0]
 
         output = self.sequence_summary(output)
         logits = self.logits_proj(output)
 
-        outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + transformer_outputs[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFXLNetForSequenceClassificationOutput(
+            loss=loss,
+            logits=logits,
+            mems=transformer_outputs.mems,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1208,7 +1409,12 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlnet-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlnet-base-cased",
+        output_type=TFXLNetForMultipleChoiceOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1223,6 +1429,7 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
         use_cache=True,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1231,24 +1438,6 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -1263,8 +1452,9 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
             use_cache = inputs[9] if len(inputs) > 9 else use_cache
             output_attentions = inputs[10] if len(inputs) > 10 else output_attentions
             output_hidden_states = inputs[11] if len(inputs) > 11 else output_hidden_states
-            labels = inputs[12] if len(inputs) > 12 else labels
+            return_dict = inputs[12] if len(inputs) > 12 else return_dict
-            assert len(inputs) <= 13, "Too many inputs."
+            labels = inputs[13] if len(inputs) > 13 else labels
+            assert len(inputs) <= 14, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -1278,10 +1468,12 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
             use_cache = inputs.get("use_cache", use_cache)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 13, "Too many inputs."
+            assert len(inputs) <= 14, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -1312,19 +1504,26 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, TFMultipleChoiceLoss):
             use_cache,
             output_attentions,
             output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
         output = transformer_outputs[0]
         logits = self.sequence_summary(output)
         logits = self.logits_proj(logits)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
-        outputs = (reshaped_logits,) + transformer_outputs[1:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, reshaped_logits)
+            output = (reshaped_logits,) + transformer_outputs[1:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (mems), (hidden states), (attentions)
+        return TFXLNetForMultipleChoiceOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            mems=transformer_outputs.mems,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1343,7 +1542,12 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
         )
 
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlnet-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlnet-base-cased",
+        output_type=TFXLNetForTokenClassificationOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1358,6 +1562,7 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
         use_cache=True,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -1365,31 +1570,12 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
-        logits (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:(batch_size, config.num_labels)`):
-            Classification scores (before SoftMax).
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[12] if len(inputs) > 12 else labels
+            labels = inputs[13] if len(inputs) > 13 else labels
-            if len(inputs) > 12:
+            if len(inputs) > 13:
-                inputs = inputs[:12]
+                inputs = inputs[:13]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -1406,19 +1592,25 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
+
         output = transformer_outputs[0]
-
         logits = self.classifier(output)
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        outputs = (logits,) + transformer_outputs[1:]  # Keep mems, hidden states, attentions if there are in it
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
 
-        if labels is not None:
+        return TFXLNetForTokenClassificationOutput(
-            loss = self.compute_loss(labels, logits)
+            loss=loss,
-            outputs = (loss,) + outputs
+            logits=logits,
-
+            mems=transformer_outputs.mems,
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -1435,7 +1627,12 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
         )
 
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xlnet-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xlnet-base-cased",
+        output_type=TFXLNetForQuestionAnsweringSimpleOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -1450,6 +1647,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
         use_cache=True,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -1463,36 +1661,13 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XLNetConfig`) and inputs:
-        loss (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
-            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
-        start_scores (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        mems (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
-            Contains pre-computed hidden-states (key and values in the attention blocks).
-            Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[12] if len(inputs) > 12 else start_positions
+            start_positions = inputs[13] if len(inputs) > 13 else start_positions
-            end_positions = inputs[13] if len(inputs) > 13 else end_positions
+            end_positions = inputs[14] if len(inputs) > 14 else end_positions
-            if len(inputs) > 12:
+            if len(inputs) > 13:
-                inputs = inputs[:12]
+                inputs = inputs[:13]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -1510,6 +1685,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
             use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -1520,17 +1696,24 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + transformer_outputs[
+        loss = None
-            1:
-        ]  # Keep mems, hidden states, attentions if there are in it
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (mems), (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFXLNetForQuestionAnsweringSimpleOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            mems=transformer_outputs.mems,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
 
 
 # @add_start_docstrings("""XLNet Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of

src/transformers/modeling_xlnet.py

@@ -711,7 +711,7 @@ class XLNetForTokenClassificationOutput(ModelOutput):
 @dataclass
 class XLNetForMultipleChoiceOutput(ModelOutput):
     """
-    Base class for outputs of multiple choice models.
+    Output type of :class:`~transformers.XLNetForMultipleChoice`.
 
     Args:
         loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
@@ -747,7 +747,7 @@ class XLNetForMultipleChoiceOutput(ModelOutput):
 @dataclass
 class XLNetForQuestionAnsweringSimpleOutput(ModelOutput):
     """
-    Base class for outputs of question answering models.
+    Output type of :class:`~transformers.XLNetForQuestionAnsweringSimple`.
 
     Args:
         loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
@@ -784,7 +784,7 @@ class XLNetForQuestionAnsweringSimpleOutput(ModelOutput):
 @dataclass
 class XLNetForQuestionAnsweringOutput(ModelOutput):
     """
-    Base class for outputs of question answering models using a :obj:`SquadHead`.
+    Output type of :class:`~transformers.XLNetForQuestionAnswering`.
 
     Args:
         loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned if both :obj:`start_positions` and :obj:`end_positions` are provided):
@@ -1227,7 +1227,6 @@ class XLNetModel(XLNetPreTrainedModel):
         # Prepare outputs, we transpose back here to shape [bsz, len, hidden_dim] (cf. beginning of forward() method)
         output = output.permute(1, 0, 2).contiguous()
 
-        # TODO Teven: fix this test to only use use_cache.
         if not use_cache:
             new_mems = None
 

templates/adding_a_new_model/modeling_tf_xxx.py

@@ -31,6 +31,14 @@ from .file_utils import (
     add_start_docstrings,
     add_start_docstrings_to_callable,
 )
+from .modeling_tf_outputs import (
+    TFBaseModelOutputWithPooling,
+    TFMaskedLMOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from .modeling_tf_utils import (
     TFMaskedLanguageModelingLoss,
     TFMultipleChoiceLoss,
@@ -46,6 +54,7 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
+_CONFIG_FOR_DOC = "XXXConfig"
 _TOKENIZER_FOR_DOC = "XxxTokenizer"
 
 ####################################################
@@ -117,35 +126,60 @@ class TFXxxMainLayer(tf.keras.layers.Layer):
         raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
 
     def call(
-        self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False
+        self,
+        inputs,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
     ):
-        # We allow three types of multi-inputs:
-        # - traditional keyword arguments in the call method
-        # - all the arguments provided as a dict in the first positional argument of call
-        # - all the arguments provided as a list/tuple (ordered) in the first positional argument of call
-        # The last two options are useful to use the tf.keras fit() method.
-
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
             attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
             token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
             position_ids = inputs[3] if len(inputs) > 3 else position_ids
             head_mask = inputs[4] if len(inputs) > 4 else head_mask
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
-        elif isinstance(inputs, dict):
+            output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
+            output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
+            assert len(inputs) <= 9, "Too many inputs."
+        elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
             token_type_ids = inputs.get("token_type_ids", token_type_ids)
             position_ids = inputs.get("position_ids", position_ids)
             head_mask = inputs.get("head_mask", head_mask)
-            assert len(inputs) <= 5, "Too many inputs."
+            inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
+            output_attentions = inputs.get("output_attentions", output_attentions)
+            output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
+            assert len(inputs) <= 9, "Too many inputs."
         else:
             input_ids = inputs
 
+        output_attentions = output_attentions if output_attentions is not None else self.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.output_hidden_states
+        return_dict = return_dict if return_dict is not None else self.return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
         if attention_mask is None:
-            attention_mask = tf.fill(shape_list(input_ids), 1)
+            attention_mask = tf.fill(input_shape, 1)
         if token_type_ids is None:
-            token_type_ids = tf.fill(shape_list(input_ids), 0)
+            token_type_ids = tf.fill(input_shape, 0)
 
         # We create a 3D attention mask from a 2D tensor mask.
         # Sizes are [batch_size, 1, 1, to_seq_length]
@@ -174,14 +208,29 @@ class TFXxxMainLayer(tf.keras.layers.Layer):
             head_mask = [None] * self.num_hidden_layers
             # head_mask = tf.constant([0] * self.num_hidden_layers)
 
-        ##################################
+        embedding_output = self.embeddings(input_ids, position_ids, token_type_ids, inputs_embeds, training=training)
-        # Replace this with your model code
+        encoder_outputs = self.encoder(
-        embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
+            embedding_output,
-        encoder_outputs = self.encoder([embedding_output, extended_attention_mask, head_mask], training=training)
+            extended_attention_mask,
-        sequence_output = encoder_outputs[0]
+            head_mask,
-        outputs = (sequence_output,) + encoder_outputs[1:]  # add hidden_states and attentions if they are here
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            training=training,
+        )
 
-        return outputs  # sequence_output, (hidden_states), (attentions)
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        if not return_dict:
+            return (sequence_output, pooled_output,) + encoder_outputs[1:]
+
+        return TFBaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
 
 
 ####################################################
@@ -274,6 +323,11 @@ XXX_INPUTS_DOCSTRING = r"""
             (if set to :obj:`False`) for evaluation.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
             If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the hidden states of all layers are returned. See ``hidden_states`` under returned tensors for more detail.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the model will return a :class:`~transformers.file_utils.ModelOutput` instead of a
+            plain tuple.
 """
 
 
@@ -287,32 +341,13 @@ class TFXxxModel(TFXxxPreTrainedModel):
         self.transformer = TFXxxMainLayer(config, name="transformer")
 
     @add_start_docstrings_to_callable(XXX_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xxx-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xxx-base-cased",
+        output_type=TFBaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(self, inputs, **kwargs):
-        r"""
-    Returns:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XxxConfig`) and inputs:
-        last_hidden_state (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`tf.Tensor` of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during XXX pretraining. This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        """
         outputs = self.transformer(inputs, **kwargs)
         return outputs
 
@@ -329,7 +364,12 @@ class TFXxxForMaskedLM(TFXxxPreTrainedModel, TFMaskedLanguageModelingLoss):
         self.mlm = TFXxxMLMHead(config, self.transformer.embeddings, name="mlm")
 
     @add_start_docstrings_to_callable(XXX_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xxx-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xxx-base-cased",
+        output_type=TFMaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -340,6 +380,7 @@ class TFXxxForMaskedLM(TFXxxPreTrainedModel, TFMaskedLanguageModelingLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -349,27 +390,12 @@ class TFXxxForMaskedLM(TFXxxPreTrainedModel, TFMaskedLanguageModelingLoss):
             Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
             in ``[0, ..., config.vocab_size]``
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XxxConfig`) and inputs:
-        prediction_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -382,19 +408,22 @@ class TFXxxForMaskedLM(TFXxxPreTrainedModel, TFMaskedLanguageModelingLoss):
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
         sequence_output = outputs[0]
         prediction_scores = self.mlm(sequence_output, training=training)
 
-        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, prediction_scores)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, prediction_scores)
+            output = (prediction_scores,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), prediction_scores, (hidden_states), (attentions)
+        return TFMaskedLMOutput(
+            loss=loss, logits=prediction_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -414,7 +443,12 @@ class TFXxxForSequenceClassification(TFXxxPreTrainedModel, TFSequenceClassificat
         )
 
     @add_start_docstrings_to_callable(XXX_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xxx-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xxx-base-cased",
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -425,6 +459,7 @@ class TFXxxForSequenceClassification(TFXxxPreTrainedModel, TFSequenceClassificat
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -434,27 +469,12 @@ class TFXxxForSequenceClassification(TFXxxPreTrainedModel, TFSequenceClassificat
             Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
             If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
             If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XxxConfig`) and inputs:
-        logits (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -467,6 +487,7 @@ class TFXxxForSequenceClassification(TFXxxPreTrainedModel, TFSequenceClassificat
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -475,13 +496,15 @@ class TFXxxForSequenceClassification(TFXxxPreTrainedModel, TFSequenceClassificat
         pooled_output = self.dropout(pooled_output, training=training)
         logits = self.classifier(pooled_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFSequenceClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -509,7 +532,12 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
     @add_start_docstrings_to_callable(XXX_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xxx-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xxx-base-cased",
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs,
@@ -520,6 +548,7 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -527,24 +556,7 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
-            of the input tensors. (see `input_ids` above)
+            of the input tensors. (see `input_ids` above)s after the attention softmax, used to compute the weighted average in the self-attention
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XxxConfig`) and inputs:
-        classification_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, num_choices)`:
-            `num_choices` is the size of the second dimension of the input tensors. (see `input_ids` above).
-
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
             heads.
         """
         if isinstance(inputs, (tuple, list)):
@@ -556,8 +568,9 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
             output_attentions = inputs[6] if len(inputs) > 6 else output_attentions
             output_hidden_states = inputs[7] if len(inputs) > 7 else output_hidden_states
-            labels = inputs[8] if len(inputs) > 8 else labels
+            return_dict = inputs[8] if len(inputs) > 8 else return_dict
-            assert len(inputs) <= 9, "Too many inputs."
+            labels = inputs[9] if len(inputs) > 9 else labels
+            assert len(inputs) <= 10, "Too many inputs."
         elif isinstance(inputs, (dict, BatchEncoding)):
             input_ids = inputs.get("input_ids")
             attention_mask = inputs.get("attention_mask", attention_mask)
@@ -567,10 +580,12 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
             inputs_embeds = inputs.get("inputs_embeds", inputs_embeds)
             output_attentions = inputs.get("output_attentions", output_attentions)
             output_hidden_states = inputs.get("output_hidden_states", output_hidden_states)
+            return_dict = inputs.get("return_dict", return_dict)
             labels = inputs.get("labels", labels)
-            assert len(inputs) <= 9, "Too many inputs."
+            assert len(inputs) <= 10, "Too many inputs."
         else:
             input_ids = inputs
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
 
         if input_ids is not None:
             num_choices = shape_list(input_ids)[1]
@@ -598,6 +613,7 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
             flat_inputs_embeds,
             output_attentions,
             output_hidden_states,
+            return_dict,
         ]
 
         outputs = self.transformer(flat_inputs, training=training)
@@ -608,13 +624,15 @@ class TFXxxForMultipleChoice(TFXxxPreTrainedModel, TFMultipleChoiceLoss):
         logits = self.classifier(pooled_output)
         reshaped_logits = tf.reshape(logits, (-1, num_choices))
 
-        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, reshaped_logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, reshaped_logits)
+            output = (reshaped_logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+        return TFMultipleChoiceModelOutput(
+            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -634,7 +652,12 @@ class TFXxxForTokenClassification(TFXxxPreTrainedModel, TFTokenClassificationLos
         )
 
     @add_start_docstrings_to_callable(XXX_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xxx-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xxx-base-cased",
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -645,6 +668,7 @@ class TFXxxForTokenClassification(TFXxxPreTrainedModel, TFTokenClassificationLos
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         labels=None,
         training=False,
     ):
@@ -652,27 +676,12 @@ class TFXxxForTokenClassification(TFXxxPreTrainedModel, TFTokenClassificationLos
         labels (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the token classification loss.
             Indices should be in ``[0, ..., config.num_labels - 1]``.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XxxConfig`) and inputs:
-        scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
-            Classification scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            labels = inputs[8] if len(inputs) > 8 else labels
+            labels = inputs[9] if len(inputs) > 9 else labels
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             labels = inputs.pop("labels", labels)
 
@@ -685,6 +694,7 @@ class TFXxxForTokenClassification(TFXxxPreTrainedModel, TFTokenClassificationLos
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -693,13 +703,15 @@ class TFXxxForTokenClassification(TFXxxPreTrainedModel, TFTokenClassificationLos
         sequence_output = self.dropout(sequence_output, training=training)
         logits = self.classifier(sequence_output)
 
-        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        loss = None if labels is None else self.compute_loss(labels, logits)
 
-        if labels is not None:
+        if not return_dict:
-            loss = self.compute_loss(labels, logits)
+            output = (logits,) + outputs[2:]
-            outputs = (loss,) + outputs
+            return ((loss,) + output) if loss is not None else output
 
-        return outputs  # (loss), logits, (hidden_states), (attentions)
+        return TFTokenClassifierOutput(
+            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
+        )
 
 
 @add_start_docstrings(
@@ -718,7 +730,12 @@ class TFXxxForQuestionAnswering(TFXxxPreTrainedModel, TFQuestionAnsweringLoss):
         )
 
     @add_start_docstrings_to_callable(XXX_INPUTS_DOCSTRING)
-    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="xxx-base-cased")
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint="xxx-base-cased",
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
     def call(
         self,
         inputs=None,
@@ -729,6 +746,7 @@ class TFXxxForQuestionAnswering(TFXxxPreTrainedModel, TFQuestionAnsweringLoss):
         inputs_embeds=None,
         output_attentions=None,
         output_hidden_states=None,
+        return_dict=None,
         start_positions=None,
         end_positions=None,
         training=False,
@@ -742,30 +760,13 @@ class TFXxxForQuestionAnswering(TFXxxPreTrainedModel, TFQuestionAnsweringLoss):
             Labels for position (index) of the end of the labelled span for computing the token classification loss.
             Positions are clamped to the length of the sequence (`sequence_length`).
             Position outside of the sequence are not taken into account for computing the loss.
-
-    Return:
-        :obj:`tuple(tf.Tensor)` comprising various elements depending on the configuration (:class:`~transformers.XxxConfig`) and inputs:
-        start_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-start scores (before SoftMax).
-        end_scores (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length,)`):
-            Span-end scores (before SoftMax).
-        hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            tuple of :obj:`tf.Tensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
         """
+        return_dict = return_dict if return_dict is not None else self.transformer.return_dict
         if isinstance(inputs, (tuple, list)):
-            start_positions = inputs[8] if len(inputs) > 8 else start_positions
+            start_positions = inputs[9] if len(inputs) > 9 else start_positions
-            end_positions = inputs[9] if len(inputs) > 9 else end_positions
+            end_positions = inputs[10] if len(inputs) > 10 else end_positions
-            if len(inputs) > 8:
+            if len(inputs) > 9:
-                inputs = inputs[:8]
+                inputs = inputs[:9]
         elif isinstance(inputs, (dict, BatchEncoding)):
             start_positions = inputs.pop("start_positions", start_positions)
             end_positions = inputs.pop("end_positions", start_positions)
@@ -779,6 +780,7 @@ class TFXxxForQuestionAnswering(TFXxxPreTrainedModel, TFQuestionAnsweringLoss):
             inputs_embeds=inputs_embeds,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
             training=training,
         )
 
@@ -789,12 +791,20 @@ class TFXxxForQuestionAnswering(TFXxxPreTrainedModel, TFQuestionAnsweringLoss):
         start_logits = tf.squeeze(start_logits, axis=-1)
         end_logits = tf.squeeze(end_logits, axis=-1)
 
-        outputs = (start_logits, end_logits,) + outputs[2:]
+        loss = None
-
         if start_positions is not None and end_positions is not None:
             labels = {"start_position": start_positions}
             labels["end_position"] = end_positions
-            loss = self.compute_loss(labels, outputs[:2])
+            loss = self.compute_loss(labels, (start_logits, end_logits))
-            outputs = (loss,) + outputs
 
-        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

templates/adding_a_new_model/tests/test_modeling_tf_xxx.py

@@ -24,9 +24,11 @@ from .utils import CACHE_DIR, require_tf, slow
 
 
 if is_tf_available():
+    import tensorflow as tf
     from transformers.modeling_tf_xxx import (
         TFXxxModel,
         TFXxxForMaskedLM,
+        TFXxxForMultipleChoice,
         TFXxxForSequenceClassification,
         TFXxxForTokenClassification,
         TFXxxForQuestionAnswering,
@@ -40,6 +42,7 @@ class TFXxxModelTest(TFModelTesterMixin, unittest.TestCase):
         (
             TFXxxModel,
             TFXxxForMaskedLM,
+            TFXxxForMultipleChoice,
             TFXxxForQuestionAnswering,
             TFXxxForSequenceClassification,
             TFXxxForTokenClassification,
@@ -128,6 +131,7 @@ class TFXxxModelTest(TFModelTesterMixin, unittest.TestCase):
                 max_position_embeddings=self.max_position_embeddings,
                 type_vocab_size=self.type_vocab_size,
                 initializer_range=self.initializer_range,
+                return_dict=True,
             )
 
             return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -137,33 +141,26 @@ class TFXxxModelTest(TFModelTesterMixin, unittest.TestCase):
         ):
             model = TFXxxModel(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            sequence_output, pooled_output = model(inputs)
+            result = model(inputs)
 
             inputs = [input_ids, input_mask]
-            sequence_output, pooled_output = model(inputs)
+            result = model(inputs)
 
-            sequence_output, pooled_output = model(input_ids)
+            result = model(input_ids)
 
-            result = {
-                "sequence_output": sequence_output.numpy(),
-                "pooled_output": pooled_output.numpy(),
-            }
             self.parent.assertListEqual(
-                list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+                list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
             )
-            self.parent.assertListEqual(list(result["pooled_output"].shape), [self.batch_size, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooler_output"].shape), [self.batch_size, self.hidden_size])
 
         def create_and_check_xxx_for_masked_lm(
             self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
         ):
             model = TFXxxForMaskedLM(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (prediction_scores,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "prediction_scores": prediction_scores.numpy(),
-            }
             self.parent.assertListEqual(
-                list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
+                list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size]
             )
 
         def create_and_check_xxx_for_sequence_classification(
@@ -172,22 +169,32 @@ class TFXxxModelTest(TFModelTesterMixin, unittest.TestCase):
             config.num_labels = self.num_labels
             model = TFXxxForSequenceClassification(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (logits,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "logits": logits.numpy(),
-            }
             self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_labels])
 
+        def create_and_check_bert_for_multiple_choice(
+            self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+        ):
+            config.num_choices = self.num_choices
+            model = TFXxxForMultipleChoice(config=config)
+            multiple_choice_inputs_ids = tf.tile(tf.expand_dims(input_ids, 1), (1, self.num_choices, 1))
+            multiple_choice_input_mask = tf.tile(tf.expand_dims(input_mask, 1), (1, self.num_choices, 1))
+            multiple_choice_token_type_ids = tf.tile(tf.expand_dims(token_type_ids, 1), (1, self.num_choices, 1))
+            inputs = {
+                "input_ids": multiple_choice_inputs_ids,
+                "attention_mask": multiple_choice_input_mask,
+                "token_type_ids": multiple_choice_token_type_ids,
+            }
+            result = model(inputs)
+            self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
+
         def create_and_check_xxx_for_token_classification(
             self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
         ):
             config.num_labels = self.num_labels
             model = TFXxxForTokenClassification(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (logits,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "logits": logits.numpy(),
-            }
             self.parent.assertListEqual(
                 list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels]
             )
@@ -197,11 +204,7 @@ class TFXxxModelTest(TFModelTesterMixin, unittest.TestCase):
         ):
             model = TFXxxForQuestionAnswering(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            start_logits, end_logits = model(inputs)
+            result = model(inputs)
-            result = {
-                "start_logits": start_logits.numpy(),
-                "end_logits": end_logits.numpy(),
-            }
             self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
             self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 

tests/test_modeling_tf_albert.py

@@ -116,6 +116,7 @@ class TFAlbertModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -129,21 +130,17 @@ class TFAlbertModelTester:
         #           'token_type_ids': token_type_ids}
         # sequence_output, pooled_output = model(**inputs)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        sequence_output, pooled_output = model(inputs)
+        result = model(inputs)
 
         inputs = [input_ids, input_mask]
-        sequence_output, pooled_output = model(inputs)
+        result = model(inputs)
 
-        sequence_output, pooled_output = model(input_ids)
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-            "pooled_output": pooled_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
-        self.parent.assertListEqual(list(result["pooled_output"].shape), [self.batch_size, self.hidden_size])
+        self.parent.assertListEqual(list(result["pooler_output"].shape), [self.batch_size, self.hidden_size])
 
     def create_and_check_albert_for_pretraining(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -151,28 +148,19 @@ class TFAlbertModelTester:
         config.num_labels = self.num_labels
         model = TFAlbertForPreTraining(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        prediction_scores, sop_scores = model(inputs)
+        result = model(inputs)
-        result = {
-            "prediction_scores": prediction_scores.numpy(),
-            "sop_scores": sop_scores.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
+            list(result["prediction_logits"].shape), [self.batch_size, self.seq_length, self.vocab_size]
         )
-        self.parent.assertListEqual(list(result["sop_scores"].shape), [self.batch_size, self.num_labels])
+        self.parent.assertListEqual(list(result["sop_logits"].shape), [self.batch_size, self.num_labels])
 
     def create_and_check_albert_for_masked_lm(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
         model = TFAlbertForMaskedLM(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (prediction_scores,) = model(inputs)
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_albert_for_sequence_classification(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -180,10 +168,7 @@ class TFAlbertModelTester:
         config.num_labels = self.num_labels
         model = TFAlbertForSequenceClassification(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_labels])
 
     def create_and_check_albert_for_question_answering(
@@ -191,11 +176,7 @@ class TFAlbertModelTester:
     ):
         model = TFAlbertForQuestionAnswering(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-        result = {
-            "start_logits": start_logits.numpy(),
-            "end_logits": end_logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 

tests/test_modeling_tf_bert.py

@@ -118,6 +118,7 @@ class TFBertModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -130,18 +131,14 @@ class TFBertModelTester:
         sequence_output, pooled_output = model(inputs)
 
         inputs = [input_ids, input_mask]
-        sequence_output, pooled_output = model(inputs)
+        result = model(inputs)
 
-        sequence_output, pooled_output = model(input_ids)
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-            "pooled_output": pooled_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
-        self.parent.assertListEqual(list(result["pooled_output"].shape), [self.batch_size, self.hidden_size])
+        self.parent.assertListEqual(list(result["pooler_output"].shape), [self.batch_size, self.hidden_size])
 
     def create_and_check_bert_lm_head(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -153,7 +150,7 @@ class TFBertModelTester:
             "attention_mask": input_mask,
             "token_type_ids": token_type_ids,
         }
-        (prediction_scores,) = model(inputs)
+        prediction_scores = model(inputs)["logits"]
         self.parent.assertListEqual(
             list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
         )
@@ -167,39 +164,27 @@ class TFBertModelTester:
             "attention_mask": input_mask,
             "token_type_ids": token_type_ids,
         }
-        (prediction_scores,) = model(inputs)
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_bert_for_next_sequence_prediction(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
         model = TFBertForNextSentencePrediction(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (seq_relationship_score,) = model(inputs)
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, 2])
-            "seq_relationship_score": seq_relationship_score.numpy(),
-        }
-        self.parent.assertListEqual(list(result["seq_relationship_score"].shape), [self.batch_size, 2])
 
     def create_and_check_bert_for_pretraining(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
         model = TFBertForPreTraining(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        prediction_scores, seq_relationship_score = model(inputs)
+        result = model(inputs)
-        result = {
-            "prediction_scores": prediction_scores.numpy(),
-            "seq_relationship_score": seq_relationship_score.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
+            list(result["prediction_logits"].shape), [self.batch_size, self.seq_length, self.vocab_size]
         )
-        self.parent.assertListEqual(list(result["seq_relationship_score"].shape), [self.batch_size, 2])
+        self.parent.assertListEqual(list(result["seq_relationship_logits"].shape), [self.batch_size, 2])
 
     def create_and_check_bert_for_sequence_classification(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -212,8 +197,7 @@ class TFBertModelTester:
             "token_type_ids": token_type_ids,
         }
 
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_labels])
 
     def create_and_check_bert_for_multiple_choice(
@@ -229,8 +213,7 @@ class TFBertModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
     def create_and_check_bert_for_token_classification(
@@ -243,10 +226,7 @@ class TFBertModelTester:
             "attention_mask": input_mask,
             "token_type_ids": token_type_ids,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels])
 
     def create_and_check_bert_for_question_answering(
@@ -259,8 +239,7 @@ class TFBertModelTester:
             "token_type_ids": token_type_ids,
         }
 
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-        result = {"start_logits": start_logits.numpy(), "end_logits": end_logits.numpy()}
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 

tests/test_modeling_tf_camembert.py

@@ -35,7 +35,7 @@ class TFCamembertModelIntegrationTest(unittest.TestCase):
             [[5, 121, 11, 660, 16, 730, 25543, 110, 83, 6]], dtype=tf.int32,
         )  # J'aime le camembert !"
 
-        output = model(input_ids)[0]
+        output = model(input_ids)["last_hidden_state"]
         expected_shape = tf.TensorShape((1, 10, 768))
         self.assertEqual(output.shape, expected_shape)
         # compare the actual values for a slice.

tests/test_modeling_tf_common.py

@@ -146,7 +146,8 @@ class TFModelTesterMixin:
                 tf.saved_model.save(model, tmpdirname)
                 model = tf.keras.models.load_model(tmpdirname)
                 outputs = model(inputs_dict)
-                hidden_states = [t.numpy() for t in outputs[-1]]
+                output = outputs[list(outputs.keys())[-1]] if isinstance(outputs, dict) else outputs[-1]
+                hidden_states = [t.numpy() for t in output]
                 self.assertEqual(len(outputs), num_out)
                 self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
                 self.assertListEqual(
@@ -177,7 +178,8 @@ class TFModelTesterMixin:
                 tf.saved_model.save(model, tmpdirname)
                 model = tf.keras.models.load_model(tmpdirname)
                 outputs = model(inputs_dict)
-                attentions = [t.numpy() for t in outputs[-1]]
+                output = outputs[list(outputs.keys())[-1]] if isinstance(outputs, dict) else outputs[-1]
+                attentions = [t.numpy() for t in output]
                 self.assertEqual(len(outputs), num_out)
                 self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
                 self.assertListEqual(
@@ -238,6 +240,8 @@ class TFModelTesterMixin:
         # Make sure we don't have nans
         if isinstance(after_outputs, tf.Tensor):
             out_1 = after_outputs.numpy()
+        elif isinstance(after_outputs, dict):
+            out_1 = after_outputs[list(after_outputs.keys())[0]]
         else:
             out_1 = after_outputs[0].numpy()
         out_2 = outputs[0].numpy()

tests/test_modeling_tf_ctrl.py

@@ -89,9 +89,10 @@ class TFCTRLModelTester(object):
             # hidden_dropout_prob=self.hidden_dropout_prob,
             # attention_probs_dropout_prob=self.attention_probs_dropout_prob,
             n_positions=self.max_position_embeddings,
-            n_ctx=self.max_position_embeddings
+            n_ctx=self.max_position_embeddings,
             # type_vocab_size=self.type_vocab_size,
-            # initializer_range=self.initializer_range
+            # initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
@@ -111,30 +112,22 @@ class TFCTRLModelTester(object):
     def create_and_check_ctrl_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
         model = TFCTRLModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
         inputs = [input_ids, None, input_mask]  # None is the input for 'past'
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
-        sequence_output = model(input_ids)[0]
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_ctrl_lm_head(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
         model = TFCTRLLMHeadModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        prediction_scores = model(inputs)[0]
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def prepare_config_and_inputs_for_common(self):
         config_and_inputs = self.prepare_config_and_inputs()

tests/test_modeling_tf_distilbert.py

@@ -89,6 +89,7 @@ class TFDistilBertModelTester:
             attention_dropout=self.attention_probs_dropout_prob,
             max_position_embeddings=self.max_position_embeddings,
             initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -99,18 +100,14 @@ class TFDistilBertModelTester:
         model = TFDistilBertModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask}
 
-        outputs = model(inputs)
+        result = model(inputs)
-        sequence_output = outputs[0]
 
         inputs = [input_ids, input_mask]
 
-        (sequence_output,) = model(inputs)
+        result = model(inputs)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_distilbert_for_masked_lm(
@@ -118,11 +115,8 @@ class TFDistilBertModelTester:
     ):
         model = TFDistilBertForMaskedLM(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask}
-        (prediction_scores,) = model(inputs)
+        result = model(inputs)
-        result = {"prediction_scores": prediction_scores.numpy()}
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_distilbert_for_question_answering(
         self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -132,8 +126,7 @@ class TFDistilBertModelTester:
             "input_ids": input_ids,
             "attention_mask": input_mask,
         }
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-        result = {"start_logits": start_logits.numpy(), "end_logits": end_logits.numpy()}
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 
@@ -143,8 +136,7 @@ class TFDistilBertModelTester:
         config.num_labels = self.num_labels
         model = TFDistilBertForSequenceClassification(config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_labels])
 
     def create_and_check_distilbert_for_multiple_choice(
@@ -158,8 +150,7 @@ class TFDistilBertModelTester:
             "input_ids": multiple_choice_inputs_ids,
             "attention_mask": multiple_choice_input_mask,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
     def create_and_check_distilbert_for_token_classification(
@@ -168,10 +159,7 @@ class TFDistilBertModelTester:
         config.num_labels = self.num_labels
         model = TFDistilBertForTokenClassification(config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels])
 
     def prepare_config_and_inputs_for_common(self):

tests/test_modeling_tf_electra.py

@@ -95,6 +95,7 @@ class TFElectraModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -104,18 +105,15 @@ class TFElectraModelTester:
     ):
         model = TFElectraModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (sequence_output,) = model(inputs)
+        result = model(inputs)
 
         inputs = [input_ids, input_mask]
-        (sequence_output,) = model(inputs)
+        result = model(inputs)
 
-        (sequence_output,) = model(input_ids)
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_electra_for_masked_lm(
@@ -123,24 +121,16 @@ class TFElectraModelTester:
     ):
         model = TFElectraForMaskedLM(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (prediction_scores,) = model(inputs)
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_electra_for_pretraining(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
         model = TFElectraForPreTraining(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (prediction_scores,) = model(inputs)
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(list(result["prediction_scores"].shape), [self.batch_size, self.seq_length])
 
     def create_and_check_electra_for_sequence_classification(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -148,10 +138,7 @@ class TFElectraModelTester:
         config.num_labels = self.num_labels
         model = TFElectraForSequenceClassification(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_labels])
 
     def create_and_check_electra_for_multiple_choice(
@@ -167,8 +154,7 @@ class TFElectraModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
     def create_and_check_electra_for_question_answering(
@@ -176,11 +162,7 @@ class TFElectraModelTester:
     ):
         model = TFElectraForQuestionAnswering(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-        result = {
-            "start_logits": start_logits.numpy(),
-            "end_logits": end_logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 
@@ -190,10 +172,7 @@ class TFElectraModelTester:
         config.num_labels = self.num_labels
         model = TFElectraForTokenClassification(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels])
 
     def prepare_config_and_inputs_for_common(self):

tests/test_modeling_tf_flaubert.py

@@ -113,6 +113,7 @@ class TFFlaubertModelTester:
             summary_type=self.summary_type,
             use_proj=self.use_proj,
             bos_token_id=self.bos_token_id,
+            return_dict=True,
         )
 
         return (
@@ -141,16 +142,12 @@ class TFFlaubertModelTester:
     ):
         model = TFFlaubertModel(config=config)
         inputs = {"input_ids": input_ids, "lengths": input_lengths, "langs": token_type_ids}
-        outputs = model(inputs)
+        result = model(inputs)
 
         inputs = [input_ids, input_mask]
-        outputs = model(inputs)
+        result = model(inputs)
-        sequence_output = outputs[0]
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_flaubert_lm_head(
@@ -168,13 +165,7 @@ class TFFlaubertModelTester:
         model = TFFlaubertWithLMHeadModel(config)
 
         inputs = {"input_ids": input_ids, "lengths": input_lengths, "langs": token_type_ids}
-        outputs = model(inputs)
+        result = model(inputs)
-
-        logits = outputs[0]
-
-        result = {
-            "logits": logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
 
@@ -194,12 +185,7 @@ class TFFlaubertModelTester:
 
         inputs = {"input_ids": input_ids, "lengths": input_lengths}
 
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-
-        result = {
-            "start_logits": start_logits.numpy(),
-            "end_logits": end_logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
@@ -220,11 +206,7 @@ class TFFlaubertModelTester:
 
         inputs = {"input_ids": input_ids, "lengths": input_lengths}
 
-        (logits,) = model(inputs)
+        result = model(inputs)
-
-        result = {
-            "logits": logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.type_sequence_label_size])
 
@@ -243,10 +225,7 @@ class TFFlaubertModelTester:
         config.num_labels = self.num_labels
         model = TFFlaubertForTokenClassification(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels])
 
     def create_and_check_flaubert_for_multiple_choice(
@@ -271,8 +250,7 @@ class TFFlaubertModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
     def prepare_config_and_inputs_for_common(self):

tests/test_modeling_tf_gpt2.py

@@ -102,6 +102,7 @@ class TFGPT2ModelTester:
             # initializer_range=self.initializer_range
             bos_token_id=self.bos_token_id,
             eos_token_id=self.eos_token_id,
+            return_dict=True,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
@@ -125,18 +126,15 @@ class TFGPT2ModelTester:
             "attention_mask": input_mask,
             "token_type_ids": token_type_ids,
         }
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
         inputs = [input_ids, None, input_mask]  # None is the input for 'past'
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
-        sequence_output = model(input_ids)[0]
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size],
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size],
         )
 
     def create_and_check_gpt2_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
@@ -150,7 +148,7 @@ class TFGPT2ModelTester:
         self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
         self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
 
-        output, past = outputs
+        output, past = outputs.to_tuple()
 
         # create hypothetical next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
@@ -160,8 +158,8 @@ class TFGPT2ModelTester:
         next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
         next_token_type_ids = tf.concat([token_type_ids, next_token_types], axis=-1)
 
-        output_from_no_past, _ = model(next_input_ids, token_type_ids=next_token_type_ids)
+        output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
-        output_from_past, _ = model(next_tokens, token_type_ids=next_token_types, past=past)
+        output_from_past = model(next_tokens, token_type_ids=next_token_types, past=past)["last_hidden_state"]
 
         # select random slice
         random_slice_idx = int(ids_tensor((1,), shape_list(output_from_past)[-1]))
@@ -183,7 +181,7 @@ class TFGPT2ModelTester:
         attn_mask = tf.concat([attn_mask_begin, attn_mask_end], axis=1)
 
         # first forward pass
-        output, past = model(input_ids, attention_mask=attn_mask)
+        output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
 
         # create hypothetical next token and extent to next_input_ids
         next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
@@ -202,8 +200,8 @@ class TFGPT2ModelTester:
         attn_mask = tf.concat([attn_mask, tf.ones((shape_list(attn_mask)[0], 1), dtype=tf.int32)], axis=1)
 
         # get two different outputs
-        output_from_no_past, _ = model(next_input_ids, attention_mask=attn_mask)
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
-        output_from_past, _ = model(next_tokens, past=past, attention_mask=attn_mask)
+        output_from_past = model(next_tokens, past=past, attention_mask=attn_mask)["last_hidden_state"]
 
         # select random slice
         random_slice_idx = int(ids_tensor((1,), shape_list(output_from_past)[-1]))
@@ -220,12 +218,9 @@ class TFGPT2ModelTester:
             "attention_mask": input_mask,
             "token_type_ids": token_type_ids,
         }
-        prediction_scores = model(inputs)[0]
+        result = model(inputs)
-        result = {
-            "prediction_scores": prediction_scores.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size],
+            list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size],
         )
 
     def create_and_check_gpt2_double_head(
@@ -243,8 +238,7 @@ class TFGPT2ModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        lm_logits, mc_logits = model(inputs)[:2]
+        result = model(inputs)
-        result = {"lm_logits": lm_logits.numpy(), "mc_logits": mc_logits.numpy()}
         self.parent.assertListEqual(
             list(result["lm_logits"].shape), [self.batch_size, self.num_choices, self.seq_length, self.vocab_size],
         )

tests/test_modeling_tf_mobilebert.py

@@ -138,6 +138,7 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
                 type_vocab_size=self.type_vocab_size,
                 initializer_range=self.initializer_range,
                 embedding_size=self.embedding_size,
+                return_dict=True,
             )
 
             return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -147,33 +148,26 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
         ):
             model = TFMobileBertModel(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            sequence_output, pooled_output = model(inputs)
+            result = model(inputs)
 
             inputs = [input_ids, input_mask]
-            sequence_output, pooled_output = model(inputs)
+            result = model(inputs)
 
-            sequence_output, pooled_output = model(input_ids)
+            result = model(input_ids)
 
-            result = {
-                "sequence_output": sequence_output.numpy(),
-                "pooled_output": pooled_output.numpy(),
-            }
             self.parent.assertListEqual(
-                list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+                list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
             )
-            self.parent.assertListEqual(list(result["pooled_output"].shape), [self.batch_size, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooler_output"].shape), [self.batch_size, self.hidden_size])
 
         def create_and_check_mobilebert_for_masked_lm(
             self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
         ):
             model = TFMobileBertForMaskedLM(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (prediction_scores,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "prediction_scores": prediction_scores.numpy(),
-            }
             self.parent.assertListEqual(
-                list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
+                list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size]
             )
 
         def create_and_check_mobilebert_for_next_sequence_prediction(
@@ -181,26 +175,19 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
         ):
             model = TFMobileBertForNextSentencePrediction(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (seq_relationship_score,) = model(inputs)
+            result = model(inputs)
-            result = {
+            self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, 2])
-                "seq_relationship_score": seq_relationship_score.numpy(),
-            }
-            self.parent.assertListEqual(list(result["seq_relationship_score"].shape), [self.batch_size, 2])
 
         def create_and_check_mobilebert_for_pretraining(
             self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
         ):
             model = TFMobileBertForPreTraining(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            prediction_scores, seq_relationship_score = model(inputs)
+            result = model(inputs)
-            result = {
-                "prediction_scores": prediction_scores.numpy(),
-                "seq_relationship_score": seq_relationship_score.numpy(),
-            }
             self.parent.assertListEqual(
-                list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
+                list(result["prediction_logits"].shape), [self.batch_size, self.seq_length, self.vocab_size]
             )
-            self.parent.assertListEqual(list(result["seq_relationship_score"].shape), [self.batch_size, 2])
+            self.parent.assertListEqual(list(result["seq_relationship_logits"].shape), [self.batch_size, 2])
 
         def create_and_check_mobilebert_for_sequence_classification(
             self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -208,10 +195,7 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
             config.num_labels = self.num_labels
             model = TFMobileBertForSequenceClassification(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (logits,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "logits": logits.numpy(),
-            }
             self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_labels])
 
         def create_and_check_mobilebert_for_multiple_choice(
@@ -227,10 +211,7 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
                 "attention_mask": multiple_choice_input_mask,
                 "token_type_ids": multiple_choice_token_type_ids,
             }
-            (logits,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "logits": logits.numpy(),
-            }
             self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
         def create_and_check_mobilebert_for_token_classification(
@@ -239,10 +220,7 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
             config.num_labels = self.num_labels
             model = TFMobileBertForTokenClassification(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            (logits,) = model(inputs)
+            result = model(inputs)
-            result = {
-                "logits": logits.numpy(),
-            }
             self.parent.assertListEqual(
                 list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels]
             )
@@ -252,11 +230,7 @@ class TFMobileBertModelTest(TFModelTesterMixin, unittest.TestCase):
         ):
             model = TFMobileBertForQuestionAnswering(config=config)
             inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-            start_logits, end_logits = model(inputs)
+            result = model(inputs)
-            result = {
-                "start_logits": start_logits.numpy(),
-                "end_logits": end_logits.numpy(),
-            }
             self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
             self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 

tests/test_modeling_tf_openai_gpt.py

@@ -94,9 +94,10 @@ class TFOpenAIGPTModelTester:
             # hidden_dropout_prob=self.hidden_dropout_prob,
             # attention_probs_dropout_prob=self.attention_probs_dropout_prob,
             n_positions=self.max_position_embeddings,
-            n_ctx=self.max_position_embeddings
+            n_ctx=self.max_position_embeddings,
             # type_vocab_size=self.type_vocab_size,
-            # initializer_range=self.initializer_range
+            # initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
@@ -116,30 +117,22 @@ class TFOpenAIGPTModelTester:
     def create_and_check_openai_gpt_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
         model = TFOpenAIGPTModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
         inputs = [input_ids, input_mask]
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
-        sequence_output = model(input_ids)[0]
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_openai_gpt_lm_head(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
         model = TFOpenAIGPTLMHeadModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        prediction_scores = model(inputs)[0]
+        result = model(inputs)
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_openai_gpt_double_head(
         self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args
@@ -156,8 +149,7 @@ class TFOpenAIGPTModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        lm_logits, mc_logits = model(inputs)[:2]
+        result = model(inputs)
-        result = {"lm_logits": lm_logits.numpy(), "mc_logits": mc_logits.numpy()}
         self.parent.assertListEqual(
             list(result["lm_logits"].shape), [self.batch_size, self.num_choices, self.seq_length, self.vocab_size]
         )

tests/test_modeling_tf_roberta.py

@@ -95,6 +95,7 @@ class TFRobertaModelTester:
             max_position_embeddings=self.max_position_embeddings,
             type_vocab_size=self.type_vocab_size,
             initializer_range=self.initializer_range,
+            return_dict=True,
         )
 
         return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -104,31 +105,23 @@ class TFRobertaModelTester:
     ):
         model = TFRobertaModel(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
         inputs = [input_ids, input_mask]
-        sequence_output = model(inputs)[0]
+        result = model(inputs)
 
-        sequence_output = model(input_ids)[0]
+        result = model(input_ids)
 
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_roberta_for_masked_lm(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
         model = TFRobertaForMaskedLM(config=config)
-        prediction_scores = model([input_ids, input_mask, token_type_ids])[0]
+        result = model([input_ids, input_mask, token_type_ids])
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_roberta_for_token_classification(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
@@ -136,10 +129,7 @@ class TFRobertaModelTester:
         config.num_labels = self.num_labels
         model = TFRobertaForTokenClassification(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels])
 
     def create_and_check_roberta_for_question_answering(
@@ -147,11 +137,7 @@ class TFRobertaModelTester:
     ):
         model = TFRobertaForQuestionAnswering(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-        result = {
-            "start_logits": start_logits.numpy(),
-            "end_logits": end_logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
 
@@ -168,10 +154,7 @@ class TFRobertaModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
     def prepare_config_and_inputs_for_common(self):

tests/test_modeling_tf_t5.py

@@ -78,6 +78,7 @@ class TFT5ModelTester:
             bos_token_id=self.pad_token_id,
             pad_token_id=self.pad_token_id,
             decoder_start_token_id=self.pad_token_id,
+            return_dict=True,
         )
 
         return (config, input_ids, input_mask, token_labels)
@@ -89,22 +90,14 @@ class TFT5ModelTester:
             "decoder_input_ids": input_ids,
             "decoder_attention_mask": input_mask,
         }
-        decoder_output, decoder_past, encoder_output = model(inputs)
+        result = model(inputs)
 
-        decoder_output, decoder_past, encoder_output = model(
+        result = model(input_ids, decoder_attention_mask=input_mask, decoder_input_ids=input_ids)
-            input_ids, decoder_attention_mask=input_mask, decoder_input_ids=input_ids
+        decoder_output = result["last_hidden_state"]
-        )
+        decoder_past = result["decoder_past_key_values"]
-        result = {
+        encoder_output = result["encoder_last_hidden_state"]
-            "encoder_output": encoder_output.numpy(),
+        self.parent.assertListEqual(list(encoder_output.shape), [self.batch_size, self.seq_length, self.hidden_size])
-            "decoder_past": decoder_past,
+        self.parent.assertListEqual(list(decoder_output.shape), [self.batch_size, self.seq_length, self.hidden_size])
-            "decoder_output": decoder_output.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["encoder_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
-        )
-        self.parent.assertListEqual(
-            list(result["decoder_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
-        )
         self.parent.assertEqual(len(decoder_past), 2)
         # decoder_past[0] should correspond to encoder output
         self.parent.assertTrue(tf.reduce_all(tf.math.equal(decoder_past[0][0], encoder_output)))
@@ -121,14 +114,9 @@ class TFT5ModelTester:
             "decoder_attention_mask": input_mask,
         }
 
-        prediction_scores, _, _ = model(inputs_dict)
+        result = model(inputs_dict)
 
-        result = {
+        self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
-            "prediction_scores": prediction_scores.numpy(),
-        }
-        self.parent.assertListEqual(
-            list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
-        )
 
     def create_and_check_t5_decoder_model_past(self, config, input_ids, decoder_input_ids, attention_mask):
         model = TFT5Model(config=config).get_decoder()

tests/test_modeling_tf_transfo_xl.py

@@ -79,6 +79,7 @@ class TFTransfoXLModelTester:
             div_val=self.div_val,
             n_layer=self.num_hidden_layers,
             eos_token_id=self.eos_token_id,
+            return_dict=True,
         )
 
         return (config, input_ids_1, input_ids_2, lm_labels)
@@ -90,11 +91,11 @@ class TFTransfoXLModelTester:
     def create_and_check_transfo_xl_model(self, config, input_ids_1, input_ids_2, lm_labels):
         model = TFTransfoXLModel(config)
 
-        hidden_states_1, mems_1 = model(input_ids_1)
+        hidden_states_1, mems_1 = model(input_ids_1).to_tuple()
 
         inputs = {"input_ids": input_ids_2, "mems": mems_1}
 
-        hidden_states_2, mems_2 = model(inputs)
+        hidden_states_2, mems_2 = model(inputs).to_tuple()
 
         result = {
             "hidden_states_1": hidden_states_1.numpy(),
@@ -121,16 +122,16 @@ class TFTransfoXLModelTester:
     def create_and_check_transfo_xl_lm_head(self, config, input_ids_1, input_ids_2, lm_labels):
         model = TFTransfoXLLMHeadModel(config)
 
-        lm_logits_1, mems_1 = model(input_ids_1)
+        lm_logits_1, mems_1 = model(input_ids_1).to_tuple()
 
         inputs = {"input_ids": input_ids_1, "labels": lm_labels}
-        _, mems_1 = model(inputs)
+        _, mems_1 = model(inputs).to_tuple()
 
-        lm_logits_2, mems_2 = model([input_ids_2, mems_1])
+        lm_logits_2, mems_2 = model([input_ids_2, mems_1]).to_tuple()
 
         inputs = {"input_ids": input_ids_1, "mems": mems_1, "labels": lm_labels}
 
-        _, mems_2 = model(inputs)
+        _, mems_2 = model(inputs).to_tuple()
 
         result = {
             "mems_1": [mem.numpy() for mem in mems_1],

tests/test_modeling_tf_xlm.py

@@ -112,6 +112,7 @@ class TFXLMModelTester:
             summary_type=self.summary_type,
             use_proj=self.use_proj,
             bos_token_id=self.bos_token_id,
+            return_dict=True,
         )
 
         return (
@@ -140,16 +141,12 @@ class TFXLMModelTester:
     ):
         model = TFXLMModel(config=config)
         inputs = {"input_ids": input_ids, "lengths": input_lengths, "langs": token_type_ids}
-        outputs = model(inputs)
+        result = model(inputs)
 
         inputs = [input_ids, input_mask]
-        outputs = model(inputs)
+        result = model(inputs)
-        sequence_output = outputs[0]
-        result = {
-            "sequence_output": sequence_output.numpy(),
-        }
         self.parent.assertListEqual(
-            list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
 
     def create_and_check_xlm_lm_head(
@@ -169,11 +166,7 @@ class TFXLMModelTester:
         inputs = {"input_ids": input_ids, "lengths": input_lengths, "langs": token_type_ids}
         outputs = model(inputs)
 
-        logits = outputs[0]
+        result = outputs
-
-        result = {
-            "logits": logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size])
 
@@ -193,12 +186,7 @@ class TFXLMModelTester:
 
         inputs = {"input_ids": input_ids, "lengths": input_lengths}
 
-        start_logits, end_logits = model(inputs)
+        result = model(inputs)
-
-        result = {
-            "start_logits": start_logits.numpy(),
-            "end_logits": end_logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
@@ -219,11 +207,7 @@ class TFXLMModelTester:
 
         inputs = {"input_ids": input_ids, "lengths": input_lengths}
 
-        (logits,) = model(inputs)
+        result = model(inputs)
-
-        result = {
-            "logits": logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.type_sequence_label_size])
 
@@ -242,10 +226,7 @@ class TFXLMModelTester:
         config.num_labels = self.num_labels
         model = TFXLMForTokenClassification(config=config)
         inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.seq_length, self.num_labels])
 
     def create_and_check_xlm_for_multiple_choice(
@@ -270,8 +251,7 @@ class TFXLMModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        (logits,) = model(inputs)
+        result = model(inputs)
-        result = {"logits": logits.numpy()}
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
 
     def prepare_config_and_inputs_for_common(self):

tests/test_modeling_tf_xlm_roberta.py

@@ -36,7 +36,7 @@ class TFFlaubertModelIntegrationTest(unittest.TestCase):
             "attention_mask": tf.convert_to_tensor([[1, 1, 1, 1, 1, 1]], dtype=tf.int32),
         }
 
-        output = model(features)[0]
+        output = model(features)["last_hidden_state"]
         expected_shape = tf.TensorShape((1, 6, 768))
         self.assertEqual(output.shape, expected_shape)
         # compare the actual values for a slice.

tests/test_modeling_tf_xlnet.py

@@ -110,6 +110,7 @@ class TFXLNetModelTester:
             bos_token_id=self.bos_token_id,
             pad_token_id=self.pad_token_id,
             eos_token_id=self.eos_token_id,
+            return_dict=True,
         )
 
         return (
@@ -147,17 +148,10 @@ class TFXLNetModelTester:
         model = TFXLNetModel(config)
 
         inputs = {"input_ids": input_ids_1, "input_mask": input_mask, "token_type_ids": segment_ids}
-
+        result = model(inputs)
-        _, _ = model(inputs)
 
         inputs = [input_ids_1, input_mask]
-
+        result = model(inputs)
-        outputs, mems_1 = model(inputs)
-
-        result = {
-            "mems_1": [mem.numpy() for mem in mems_1],
-            "outputs": outputs.numpy(),
-        }
 
         config.mem_len = 0
         model = TFXLNetModel(config)
@@ -165,10 +159,10 @@ class TFXLNetModelTester:
         self.parent.assertEqual(len(no_mems_outputs), 1)
 
         self.parent.assertListEqual(
-            list(result["outputs"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size]
         )
         self.parent.assertListEqual(
-            list(list(mem.shape) for mem in result["mems_1"]),
+            list(list(mem.shape) for mem in result["mems"]),
             [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
         )
 
@@ -189,16 +183,13 @@ class TFXLNetModelTester:
         model = TFXLNetLMHeadModel(config)
 
         inputs_1 = {"input_ids": input_ids_1, "token_type_ids": segment_ids}
-
+        all_logits_1, mems_1 = model(inputs_1).to_tuple()
-        all_logits_1, mems_1 = model(inputs_1)
 
         inputs_2 = {"input_ids": input_ids_2, "mems": mems_1, "token_type_ids": segment_ids}
-
+        all_logits_2, mems_2 = model(inputs_2).to_tuple()
-        all_logits_2, mems_2 = model(inputs_2)
 
         inputs_3 = {"input_ids": input_ids_q, "perm_mask": perm_mask, "target_mapping": target_mapping}
-
+        logits, _ = model(inputs_3).to_tuple()
-        logits, _ = model(inputs_3)
 
         result = {
             "mems_1": [mem.numpy() for mem in mems_1],
@@ -240,13 +231,7 @@ class TFXLNetModelTester:
         model = TFXLNetForQuestionAnsweringSimple(config)
 
         inputs = {"input_ids": input_ids_1, "attention_mask": input_mask, "token_type_ids": segment_ids}
-        start_logits, end_logits, mems = model(inputs)
+        result = model(inputs)
-
-        result = {
-            "start_logits": start_logits.numpy(),
-            "end_logits": end_logits.numpy(),
-            "mems": [m.numpy() for m in mems],
-        }
 
         self.parent.assertListEqual(list(result["start_logits"].shape), [self.batch_size, self.seq_length])
         self.parent.assertListEqual(list(result["end_logits"].shape), [self.batch_size, self.seq_length])
@@ -271,16 +256,11 @@ class TFXLNetModelTester:
     ):
         model = TFXLNetForSequenceClassification(config)
 
-        logits, mems_1 = model(input_ids_1)
+        result = model(input_ids_1)
-
-        result = {
-            "mems_1": [mem.numpy() for mem in mems_1],
-            "logits": logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.type_sequence_label_size])
         self.parent.assertListEqual(
-            list(list(mem.shape) for mem in result["mems_1"]),
+            list(list(mem.shape) for mem in result["mems"]),
             [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
         )
 
@@ -305,16 +285,12 @@ class TFXLNetModelTester:
             "attention_mask": input_mask,
             # 'token_type_ids': token_type_ids
         }
-        logits, mems_1 = model(inputs)
+        result = model(inputs)
-        result = {
-            "mems_1": [mem.numpy() for mem in mems_1],
-            "logits": logits.numpy(),
-        }
         self.parent.assertListEqual(
             list(result["logits"].shape), [self.batch_size, self.seq_length, config.num_labels]
         )
         self.parent.assertListEqual(
-            list(list(mem.shape) for mem in result["mems_1"]),
+            list(list(mem.shape) for mem in result["mems"]),
             [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
         )
 
@@ -342,15 +318,11 @@ class TFXLNetModelTester:
             "attention_mask": multiple_choice_input_mask,
             "token_type_ids": multiple_choice_token_type_ids,
         }
-        (logits, mems_1) = model(inputs)
+        result = model(inputs)
-        result = {
-            "mems_1": [mem.numpy() for mem in mems_1],
-            "logits": logits.numpy(),
-        }
 
         self.parent.assertListEqual(list(result["logits"].shape), [self.batch_size, self.num_choices])
         self.parent.assertListEqual(
-            list(list(mem.shape) for mem in result["mems_1"]),
+            list(list(mem.shape) for mem in result["mems"]),
             [[self.seq_length, self.batch_size * self.num_choices, self.hidden_size]] * self.num_hidden_layers,
         )