PyTorch XLM

docs/source/model_doc/xlm.rst

@@ -1,75 +1,88 @@
 XLM
 ----------------------------------------------------
 
-``XLMConfig``
+The XLM model was proposed in `Cross-lingual Language Model Pretraining`_
+by Guillaume Lample*, Alexis Conneau*. It's a transformer pre-trained using one of the following objectives:
+
+    - a causal language modeling (CLM) objective (next token prediction),
+    - a masked language modeling (MLM) objective (Bert-like), or
+    - a Translation Language Modeling (TLM) object (extension of Bert's MLM to multiple language inputs)
+
+Original code can be found `here <https://github.com/facebookresearch/XLM>`_.
+
+This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+
+XLMConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMConfig
     :members:
 
-``XLMTokenizer``
+XLMTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMTokenizer
     :members:
 
-``XLMModel``
+XLMModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMModel
     :members:
 
 
-``XLMWithLMHeadModel``
+XLMWithLMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMWithLMHeadModel
     :members:
 
 
-``XLMForSequenceClassification``
+XLMForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMForSequenceClassification
     :members:
 
 
-``XLMForQuestionAnsweringSimple``
+XLMForQuestionAnsweringSimple
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMForQuestionAnsweringSimple
     :members:
 
 
-``XLMForQuestionAnswering``
+XLMForQuestionAnswering
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.XLMForQuestionAnswering
     :members:
 
 
-``TFXLMModel``
+TFXLMModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.TFXLMModel
     :members:
 
 
-``TFXLMWithLMHeadModel``
+TFXLMWithLMHeadModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.TFXLMWithLMHeadModel
     :members:
 
 
-``TFXLMForSequenceClassification``
+TFXLMForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.TFXLMForSequenceClassification
     :members:
 
 
-``TFXLMForQuestionAnsweringSimple``
+TFXLMForQuestionAnsweringSimple
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.TFXLMForQuestionAnsweringSimple

src/transformers/modeling_xlm.py

@@ -27,7 +27,7 @@ from torch.nn import CrossEntropyLoss, MSELoss
 from torch.nn import functional as F
 
 from .configuration_xlm import XLMConfig
-from .file_utils import add_start_docstrings
+from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
 from .modeling_utils import PreTrainedModel, SequenceSummary, SQuADHead, prune_linear_layer
 
 
@@ -249,27 +249,7 @@ class XLMPreTrainedModel(PreTrainedModel):
             module.weight.data.fill_(1.0)
 
 
-XLM_START_DOCSTRING = r"""    The XLM model was proposed in
+XLM_START_DOCSTRING = r"""
-    `Cross-lingual Language Model Pretraining`_
-    by Guillaume Lample*, Alexis Conneau*. It's a transformer pre-trained using one of the following objectives:
-
-        - a causal language modeling (CLM) objective (next token prediction),
-        - a masked language modeling (MLM) objective (Bert-like), or
-        - a Translation Language Modeling (TLM) object (extension of Bert's MLM to multiple language inputs)
-
-    Original code can be found `here`_.
-
-    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
-    refer to the PyTorch documentation for all matter related to general usage and behavior.
-
-    .. _`Cross-lingual Language Model Pretraining`:
-        https://arxiv.org/abs/1901.07291
-
-    .. _`torch.nn.Module`:
-        https://pytorch.org/docs/stable/nn.html#module
-
-    .. _`here`:
-        https://github.com/facebookresearch/XLM
 
     Parameters:
         config (:class:`~transformers.XLMConfig`): Model configuration class with all the parameters of the model.
@@ -278,48 +258,55 @@ XLM_START_DOCSTRING = r"""    The XLM model was proposed in
 """
 
 XLM_INPUTS_DOCSTRING = r"""
-    Inputs:
+    Args:
-        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
             Indices of input sequence tokens in the vocabulary. 
             
-            XLM is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            Indices can be obtained using :class:`transformers.BertTokenizer`.
-            the right rather than the left.
-
-            Indices can be obtained using :class:`transformers.XLMTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
-            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+            :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
-        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **langs**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        langs (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`): 
             A parallel sequence of tokens to be used to indicate the language of each token in the input.
             Indices are languages ids which can be obtained from the language names by using two conversion mappings
             provided in the configuration of the model (only provided for multilingual models).
             More precisely, the `language name -> language id` mapping is in `model.config.lang2id` (dict str -> int) and
             the `language id -> language name` mapping is `model.config.id2lang` (dict int -> str).
-        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            
-            A parallel sequence of tokens (can be used to indicate various portions of the inputs).
+            See usage examples detailed in the `multilingual documentation <https://huggingface.co/transformers/multilingual.html>`__.
-            The embeddings from these tokens will be summed with the respective token embeddings.
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`): 
-            Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
+            Segment token indices to indicate first and second portions of the inputs.
-        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+            
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
-        **lengths**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            
+            `What are position IDs? <../glossary.html#position-ids>`_
+        lengths (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`): 
             Length of each sentence that can be used to avoid performing attention on padding token indices.
             You can also use `attention_mask` for the same result (see above), kept here for compatbility.
             Indices selected in ``[0, ..., input_ids.size(-1)]``:
-        **cache**:
+        cache (:obj:`Dict[str, torch.FloatTensor]`, `optional`, defaults to :obj:`None`):
             dictionary with ``torch.FloatTensor`` that contains pre-computed
             hidden-states (key and values in the attention blocks) as computed by the model
             (see `cache` output below). Can be used to speed up sequential decoding.
             The dictionary object will be modified in-place during the forward pass to add newly computed hidden-states.
-        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
-            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+            :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+        input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
-            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
 """
@@ -328,30 +315,8 @@ XLM_INPUTS_DOCSTRING = r"""
 @add_start_docstrings(
     "The bare XLM Model transformer outputting raw hidden-states without any specific head on top.",
     XLM_START_DOCSTRING,
-    XLM_INPUTS_DOCSTRING,
 )
 class XLMModel(XLMPreTrainedModel):
-    r"""
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
-            Sequence of hidden-states at the last layer of the model.
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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::
-
-        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        model = XLMModel.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids)
-        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
-
-    """
 
     def __init__(self, config):  # , dico, is_encoder, with_output):
         super().__init__(config)
@@ -437,6 +402,7 @@ class XLMModel(XLMPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.attentions[layer].prune_heads(heads)
 
+    @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -448,7 +414,33 @@ class XLMModel(XLMPreTrainedModel):
         cache=None,
         head_mask=None,
         inputs_embeds=None,
-    ):  # removed: src_enc=None, src_len=None
+    ):
+        r"""
+    Return:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:obj:`~transformers.GPT2Config`) and inputs:
+        last_hidden_state (:obj:`torch.FloatTensor` 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(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (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(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (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::
+
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMModel.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+        """
         if input_ids is not None:
             bs, slen = input_ids.size()
         else:
@@ -626,39 +618,8 @@ class XLMPredLayer(nn.Module):
     """The XLM Model transformer with a language modeling head on top
     (linear layer with weights tied to the input embeddings). """,
     XLM_START_DOCSTRING,
-    XLM_INPUTS_DOCSTRING,
 )
 class XLMWithLMHeadModel(XLMPreTrainedModel):
-    r"""
-        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
-            Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
-            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
-            All labels set to ``-100`` are ignored (masked), the loss is only
-            computed for labels in ``[0, ..., config.vocab_size]``
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Language modeling loss.
-        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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::
-
-        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        model = XLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids)
-        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
-
-    """
 
     def __init__(self, config):
         super().__init__(config)
@@ -683,6 +644,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
             langs = None
         return {"input_ids": input_ids, "langs": langs}
 
+    @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -696,6 +658,41 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
         inputs_embeds=None,
         labels=None,
     ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Labels for language modeling.
+            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
+            All labels set to ``-100`` are ignored (masked), the loss is only
+            computed for labels in ``[0, ..., config.vocab_size]``
+
+    Return:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:obj:`~transformers.GPT2Config`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when ``labels`` is provided)
+            Language modeling loss.
+        prediction_scores (:obj:`torch.FloatTensor` 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(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (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(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (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::
+
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+        """
         transformer_outputs = self.transformer(
             input_ids,
             attention_mask=attention_mask,
@@ -719,39 +716,8 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
     """XLM Model with a sequence classification/regression head on top (a linear layer on top of
     the pooled output) e.g. for GLUE tasks. """,
     XLM_START_DOCSTRING,
-    XLM_INPUTS_DOCSTRING,
 )
 class XLMForSequenceClassification(XLMPreTrainedModel):
-    r"""
-        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels for computing the sequence classification/regression loss.
-            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).
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Classification (or regression if config.num_labels==1) loss.
-        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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::
-
-        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        model = XLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids, labels=labels)
-        loss, logits = outputs[:2]
-
-    """
 
     def __init__(self, config):
         super().__init__(config)
@@ -762,6 +728,7 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
 
         self.init_weights()
 
+    @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -775,6 +742,41 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
         inputs_embeds=None,
         labels=None,
     ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the sequence classification/regression loss.
+            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).
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (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(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (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::
+
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+        """
         transformer_outputs = self.transformer(
             input_ids,
             attention_mask=attention_mask,
@@ -809,51 +811,8 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
     """XLM Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
     the hidden-states output to compute `span start logits` and `span end logits`). """,
     XLM_START_DOCSTRING,
-    XLM_INPUTS_DOCSTRING,
 )
 class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
-    r"""
-        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels for position (index) of the start 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.
-        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            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.
-        **is_impossible**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels whether a question has an answer or no answer (SQuAD 2.0)
-        **cls_index**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels for position (index) of the classification token to use as input for computing plausibility of the answer.
-        **p_mask**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
-            Optional mask of tokens which can't be in answers (e.g. [CLS], [PAD], ...)
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
-        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
-            Span-start scores (before SoftMax).
-        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
-            Span-end scores (before SoftMax).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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::
-
-        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        model = XLMForQuestionAnsweringSimple.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        start_positions = torch.tensor([1])
-        end_positions = torch.tensor([3])
-        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
-        loss, start_scores, end_scores = outputs[:2]
-
-    """
 
     def __init__(self, config):
         super().__init__(config)
@@ -863,6 +822,7 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
 
         self.init_weights()
 
+    @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -877,6 +837,47 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
         start_positions=None,
         end_positions=None,
     ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the start 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.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            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(torch.FloatTensor)` comprising various elements depending on the configuration (config) and inputs:
+        loss (:obj:`torch.FloatTensor` 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:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-start scores (before SoftMax).
+        end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-end scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (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(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (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::
+
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMForQuestionAnsweringSimple.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        start_positions = torch.tensor([1])
+        end_positions = torch.tensor([3])
+        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
+        loss, start_scores, end_scores = outputs[:2]
+
+        """
         transformer_outputs = self.transformer(
             input_ids,
             attention_mask=attention_mask,
@@ -926,51 +927,8 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
     """XLM Model with a beam-search span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
     the hidden-states output to compute `span start logits` and `span end logits`). """,
     XLM_START_DOCSTRING,
-    XLM_INPUTS_DOCSTRING,
 )
 class XLMForQuestionAnswering(XLMPreTrainedModel):
-    r"""
-        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels for position (index) of the start 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.
-        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            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.
-        **is_impossible**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels whether a question has an answer or no answer (SQuAD 2.0)
-        **cls_index**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
-            Labels for position (index) of the classification token to use as input for computing plausibility of the answer.
-        **p_mask**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
-            Optional mask of tokens which can't be in answers (e.g. [CLS], [PAD], ...)
-
-    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
-        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
-            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
-        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
-            Span-start scores (before SoftMax).
-        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
-            Span-end scores (before SoftMax).
-        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
-            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
-            of shape ``(batch_size, sequence_length, hidden_size)``:
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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::
-
-        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        model = XLMForQuestionAnswering.from_pretrained('xlm-mlm-en-2048')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
-        start_positions = torch.tensor([1])
-        end_positions = torch.tensor([3])
-        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
-        loss, start_scores, end_scores = outputs[:2]
-
-    """
 
     def __init__(self, config):
         super().__init__(config)
@@ -980,6 +938,7 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
 
         self.init_weights()
 
+    @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -997,6 +956,60 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
         cls_index=None,
         p_mask=None,
     ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the start 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.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            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.
+        is_impossible (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`, defaults to :obj:`None`):
+            Labels whether a question has an answer or no answer (SQuAD 2.0)
+        cls_index (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the classification token to use as input for computing plausibility of the answer.
+        p_mask (``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``, `optional`, defaults to :obj:`None`):
+            Optional mask of tokens which can't be in answers (e.g. [CLS], [PAD], ...).
+            1.0 means token should be masked. 0.0 mean token is not masked.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (config) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned if both :obj:`start_positions` and :obj:`end_positions` are provided):
+            Classification loss as the sum of start token, end token (and is_impossible if provided) classification losses.
+        start_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
+        start_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Indices for the top config.start_n_top start token possibilities (beam-search).
+        end_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        end_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        cls_logits (``torch.FloatTensor`` of shape ``(batch_size,)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Log probabilities for the ``is_impossible`` label of the answers.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (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(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (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::
+
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMForQuestionAnswering.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        start_positions = torch.tensor([1])
+        end_positions = torch.tensor([3])
+        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
+        loss, start_scores, end_scores = outputs[:2]
+
+        """
         transformer_outputs = self.transformer(
             input_ids,
             attention_mask=attention_mask,