SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[EncoderDecoder] Add encoder-decoder for roberta/ vanilla longformer (#6411)

Browse Source 
* add encoder-decoder for roberta

* fix headmask

* apply Sylvains suggestions

* fix typo

* Apply suggestions from code review
This commit is contained in:
Patrick von Platen
2020-08-12 18:23:30 +02:00
committed by GitHub
parent fd3de2000f
commit 0735def8e1
10 changed files with 671 additions and 357 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
src/transformers/__init__.py
View File

@@ -302,6 +302,7 @@ if is_torch_available():
from .tokenization_marian import MarianTokenizer
from .modeling_roberta import (
RobertaForMaskedLM,
RobertaForCausalLM,
RobertaModel,
RobertaForSequenceClassification,
RobertaForMultipleChoice,

2
src/transformers/modeling_auto.py
View File

@@ -135,6 +135,7 @@ from .modeling_reformer import (
)
from .modeling_retribert import RetriBertModel
from .modeling_roberta import (
RobertaForCausalLM,
RobertaForMaskedLM,
RobertaForMultipleChoice,
RobertaForQuestionAnswering,
@@ -250,6 +251,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
[
(RobertaConfig, RobertaForCausalLM),
(BertConfig, BertLMHeadModel),
(OpenAIGPTConfig, OpenAIGPTLMHeadModel),
(GPT2Config, GPT2LMHeadModel),

39
src/transformers/modeling_bert.py
View File

@@ -683,14 +683,6 @@ BERT_INPUTS_DOCSTRING = r"""
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.
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
if the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask
is used in the cross-attention if the model is configured as a decoder.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
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`):
@@ -769,6 +761,16 @@ class BertModel(BertPreTrainedModel):
output_hidden_states=None,
return_dict=None,
):
r"""
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
if the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask
is used in the cross-attention if the model is configured as a decoder.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
"""
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
@@ -956,7 +958,7 @@ class BertLMHeadModel(BertPreTrainedModel):
super().__init__(config)
if not config.is_decoder:
logger.info("If you want to use `BertLMHeadModel` as a standalone, add `is_decoder=True.`")
logger.warning("If you want to use `BertLMHeadModel` as a standalone, add `is_decoder=True.`")
self.bert = BertModel(config)
self.cls = BertOnlyMLMHead(config)
@@ -976,22 +978,27 @@ class BertLMHeadModel(BertPreTrainedModel):
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
**kwargs
):
r"""
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
if the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask
is used in the cross-attention if the model is configured as a decoder.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Labels for computing the left-to-right language modeling loss (next word prediction).
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]``
kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
Used to hide legacy arguments that have been deprecated.
Returns:
@@ -1061,8 +1068,8 @@ class BertForMaskedLM(BertPreTrainedModel):
super().__init__(config)
if config.is_decoder:
logger.info(
"If you want to use `TFBertForMaskedLM` make sure `config.is_decoder=False` for "
logger.warning(
"If you want to use `BertForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
@@ -1089,9 +1096,9 @@ class BertForMaskedLM(BertPreTrainedModel):
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,

12
src/transformers/modeling_encoder_decoder.py
View File

@@ -191,11 +191,9 @@ class EncoderDecoderModel(PreTrainedModel):
input_ids=None,
inputs_embeds=None,
attention_mask=None,
head_mask=None,
encoder_outputs=None,
decoder_input_ids=None,
decoder_attention_mask=None,
decoder_head_mask=None,
decoder_inputs_embeds=None,
labels=None,
**kwargs,
@@ -216,10 +214,6 @@ class EncoderDecoderModel(PreTrainedModel):
Mask to avoid performing attention on padding token indices for the encoder.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
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 for the encoder.
Mask values selected in ``[0, 1]``:
``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
encoder_outputs (:obj:`tuple(tuple(torch.FloatTensor)`, `optional`, defaults to :obj:`None`):
Tuple consists of (`last_hidden_state`, `optional`: `hidden_states`, `optional`: `attentions`)
`last_hidden_state` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`) is a sequence of hidden-states at the output of the last layer of the encoder.
@@ -231,10 +225,6 @@ class EncoderDecoderModel(PreTrainedModel):
:func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
decoder_attention_mask (:obj:`torch.BoolTensor` of shape :obj:`(batch_size, tgt_seq_len)`, `optional`, defaults to :obj:`None`):
Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.
decoder_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 for the decoder.
Mask values selected in ``[0, 1]``:
``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
decoder_inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, target_sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
Optionally, instead of passing :obj:`decoder_input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors
@@ -279,7 +269,6 @@ class EncoderDecoderModel(PreTrainedModel):
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
head_mask=head_mask,
return_dict=False,
**kwargs_encoder,
)
@@ -293,7 +282,6 @@ class EncoderDecoderModel(PreTrainedModel):
attention_mask=decoder_attention_mask,
encoder_hidden_states=hidden_states,
encoder_attention_mask=attention_mask,
head_mask=decoder_head_mask,
labels=labels,
return_dict=False,
**kwargs_decoder,

133
src/transformers/modeling_roberta.py
View File

@@ -24,9 +24,15 @@ import torch.nn as nn
from torch.nn import CrossEntropyLoss, MSELoss
from .configuration_roberta import RobertaConfig
from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
from .file_utils import (
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_callable,
replace_return_docstrings,
)
from .modeling_bert import BertEmbeddings, BertLayerNorm, BertModel, BertPreTrainedModel, gelu
from .modeling_outputs import (
CausalLMOutput,
MaskedLMOutput,
MultipleChoiceModelOutput,
QuestionAnsweringModelOutput,
@@ -175,6 +181,121 @@ class RobertaModel(BertModel):
self.embeddings.word_embeddings = value
@add_start_docstrings(
"""RoBERTa Model with a `language modeling` head on top for CLM fine-tuning. """, ROBERTA_START_DOCSTRING
)
class RobertaForCausalLM(BertPreTrainedModel):
config_class = RobertaConfig
base_model_prefix = "roberta"
def __init__(self, config):
super().__init__(config)
if not config.is_decoder:
logger.warning("If you want to use `RobertaLMHeadModel` as a standalone, add `is_decoder=True.`")
self.roberta = RobertaModel(config)
self.lm_head = RobertaLMHead(config)
self.init_weights()
def get_output_embeddings(self):
return self.lm_head.decoder
@add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@replace_return_docstrings(output_type=CausalLMOutput, config_class=_CONFIG_FOR_DOC)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
if the model is configured as a decoder.
encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask
is used in the cross-attention if the model is configured as a decoder.
Mask values selected in ``[0, 1]``:
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
Labels for computing the left-to-right language modeling loss (next word prediction).
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:
Example::
>>> from transformers import RobertaTokenizer, RobertaLMHeadModel, RobertaConfig
>>> import torch
>>> tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
>>> config = RobertaConfig.from_pretrained("roberta-base")
>>> config.is_decoder = True
>>> model = RobertaLMHeadModel.from_pretrained('roberta-base', config=config, return_dict=True)
>>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
>>> outputs = model(**inputs)
>>> prediction_logits = outputs.logits
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.roberta(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
prediction_scores = self.lm_head(sequence_output)
lm_loss = None
if labels is not None:
# we are doing next-token prediction; shift prediction scores and input ids by one
shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
labels = labels[:, 1:].contiguous()
loss_fct = CrossEntropyLoss()
lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
if not return_dict:
output = (prediction_scores,) + outputs[2:]
return ((lm_loss,) + output) if lm_loss is not None else output
return CausalLMOutput(
loss=lm_loss, logits=prediction_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
)
def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
input_shape = input_ids.shape
# if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
if attention_mask is None:
attention_mask = input_ids.new_ones(input_shape)
return {"input_ids": input_ids, "attention_mask": attention_mask}
@add_start_docstrings("""RoBERTa Model with a `language modeling` head on top. """, ROBERTA_START_DOCSTRING)
class RobertaForMaskedLM(BertPreTrainedModel):
config_class = RobertaConfig
@@ -183,6 +304,12 @@ class RobertaForMaskedLM(BertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
if config.is_decoder:
logger.warning(
"If you want to use `RobertaForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
self.roberta = RobertaModel(config)
self.lm_head = RobertaLMHead(config)
@@ -206,6 +333,8 @@ class RobertaForMaskedLM(BertPreTrainedModel):
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
@@ -237,6 +366,8 @@ class RobertaForMaskedLM(BertPreTrainedModel):
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,

4
src/transformers/modeling_tf_bert.py
View File

@@ -862,7 +862,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel, TFMaskedLanguageModelingLoss):
super().__init__(config, *inputs, **kwargs)
if config.is_decoder:
logger.info(
logger.warning(
"If you want to use `TFBertForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
@@ -941,7 +941,7 @@ class TFBertLMHeadModel(TFBertPreTrainedModel, TFCausalLanguageModelingLoss):
super().__init__(config, *inputs, **kwargs)
if not config.is_decoder:
logger.info("If you want to use `TFBertLMHeadModel` as a standalone, add `is_decoder=True.`")
logger.warning("If you want to use `TFBertLMHeadModel` as a standalone, add `is_decoder=True.`")
self.bert = TFBertMainLayer(config, name="bert")
self.mlm = TFBertMLMHead(config, self.bert.embeddings, name="mlm___cls")