modeling_xxx.py 31.2 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
# coding=utf-8
# Copyright 2018 XXX Authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch XXX model. """

####################################################
# In this template, replace all the XXX (various casings) with your model name
####################################################


import logging
import os

import torch
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss

from .configuration_xxx import XxxConfig
30
31
32
33
34
35
36
37
38
from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
from .modeling_outputs import (
    BaseModelOutputWithPooling,
    MaskedLMOutput,
    MultipleChoiceModelOutput,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
39
from .modeling_utils import PreTrainedModel
Aymeric Augustin's avatar
Aymeric Augustin committed
40

thomwolf's avatar
thomwolf committed
41
42
43

logger = logging.getLogger(__name__)

44
45
46
_CONFIG_FOR_DOC = "XXXConfig"
_TOKENIZER_FOR_DOC = "XXXTokenizer"

thomwolf's avatar
thomwolf committed
47
####################################################
48
49
# This list contrains shortcut names for some of
# the pretrained weights provided with the models
thomwolf's avatar
thomwolf committed
50
####################################################
51
52
53
54
XXX_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "xxx-base-uncased",
    "xxx-large-uncased",
]
thomwolf's avatar
thomwolf committed
55

56

thomwolf's avatar
thomwolf committed
57
58
59
60
61
62
63
64
65
66
67
68
####################################################
# This is a conversion method from TF 1.0 to PyTorch
# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
####################################################
def load_tf_weights_in_xxx(model, config, tf_checkpoint_path):
    """ Load tf checkpoints in a pytorch model.
    """
    try:
        import re
        import numpy as np
        import tensorflow as tf
    except ImportError:
69
70
71
72
        logger.error(
            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions."
        )
thomwolf's avatar
thomwolf committed
73
74
75
76
77
78
79
80
81
82
83
84
85
86
        raise
    tf_path = os.path.abspath(tf_checkpoint_path)
    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
        logger.info("Loading TF weight {} with shape {}".format(name, shape))
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array)

    for name, array in zip(names, arrays):
87
        name = name.split("/")
thomwolf's avatar
thomwolf committed
88
89
        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
        # which are not required for using pretrained model
90
91
92
93
        if any(
            n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
            for n in name
        ):
thomwolf's avatar
thomwolf committed
94
95
96
97
            logger.info("Skipping {}".format("/".join(name)))
            continue
        pointer = model
        for m_name in name:
98
            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
99
                scope_names = re.split(r"_(\d+)", m_name)
thomwolf's avatar
thomwolf committed
100
            else:
101
102
                scope_names = [m_name]
            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
103
                pointer = getattr(pointer, "weight")
104
            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
105
                pointer = getattr(pointer, "bias")
106
            elif scope_names[0] == "output_weights":
107
                pointer = getattr(pointer, "weight")
108
            elif scope_names[0] == "squad":
109
                pointer = getattr(pointer, "classifier")
thomwolf's avatar
thomwolf committed
110
111
            else:
                try:
112
                    pointer = getattr(pointer, scope_names[0])
thomwolf's avatar
thomwolf committed
113
114
115
                except AttributeError:
                    logger.info("Skipping {}".format("/".join(name)))
                    continue
116
117
            if len(scope_names) >= 2:
                num = int(scope_names[1])
thomwolf's avatar
thomwolf committed
118
                pointer = pointer[num]
119
120
121
        if m_name[-11:] == "_embeddings":
            pointer = getattr(pointer, "weight")
        elif m_name == "kernel":
thomwolf's avatar
thomwolf committed
122
123
124
125
126
127
128
129
130
131
132
133
134
135
            array = np.transpose(array)
        try:
            assert pointer.shape == array.shape
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        logger.info("Initialize PyTorch weight {}".format(name))
        pointer.data = torch.from_numpy(array)
    return model


####################################################
# PyTorch Models are constructed by sub-classing
# - torch.nn.Module for the layers and
Julien Chaumond's avatar
Julien Chaumond committed
136
# - PreTrainedModel for the models (itself a sub-class of torch.nn.Module)
thomwolf's avatar
thomwolf committed
137
138
139
140
141
142
143
144
####################################################

####################################################
# Here is an example of typical layer in a PyTorch model of the library
# The classes are usually identical to the TF 2.0 ones without the 'TF' prefix.
#
# See the conversion methods in modeling_tf_pytorch_utils.py for more details
####################################################
145
146
147
148
149
150
151
152

XxxAttention = nn.Module

XxxIntermediate = nn.Module

XxxOutput = nn.Module


thomwolf's avatar
thomwolf committed
153
154
class XxxLayer(nn.Module):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
155
        super().__init__()
thomwolf's avatar
thomwolf committed
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
        self.attention = XxxAttention(config)
        self.intermediate = XxxIntermediate(config)
        self.output = XxxOutput(config)

    def forward(self, hidden_states, attention_mask=None, head_mask=None):
        attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
        attention_output = attention_outputs[0]
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        outputs = (layer_output,) + attention_outputs[1:]  # add attentions if we output them
        return outputs


####################################################
# PreTrainedModel is a sub-class of torch.nn.Module
# which take care of loading and saving pretrained weights
# and various common utilities.
#
# Here you just need to specify a few (self-explanatory)
# pointers for your model and the weights initialization
# method if its not fully covered by PreTrainedModel's default method
####################################################
178
179
180
181
182
183
184
185
186
187

XxxLayerNorm = torch.nn.LayerNorm

XxxEmbeddings = nn.Module

XxxEncoder = nn.Module

XxxPooler = nn.Module


thomwolf's avatar
thomwolf committed
188
189
class XxxPreTrainedModel(PreTrainedModel):
    """ An abstract class to handle weights initialization and
190
        a simple interface for downloading and loading pretrained models.
thomwolf's avatar
thomwolf committed
191
    """
192

thomwolf's avatar
thomwolf committed
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
    config_class = XxxConfig
    load_tf_weights = load_tf_weights_in_xxx
    base_model_prefix = "transformer"

    def _init_weights(self, module):
        """ Initialize the weights """
        if isinstance(module, (nn.Linear, nn.Embedding)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, XxxLayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        if isinstance(module, nn.Linear) and module.bias is not None:
            module.bias.data.zero_()


XXX_START_DOCSTRING = r"""    The XXX model was proposed in
211
212
    `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding
    <https://arxiv.org/abs/1810.04805>`__ by....
thomwolf's avatar
thomwolf committed
213

214
215
216
    This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#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.
thomwolf's avatar
thomwolf committed
217
218

    Parameters:
219
        config (:class:`~transformers.XxxConfig`): Model configuration class with all the parameters of the model.
thomwolf's avatar
thomwolf committed
220
221
222
223
224
225
            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.
"""

XXX_INPUTS_DOCSTRING = r"""
    Inputs:
226
        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
thomwolf's avatar
thomwolf committed
227
228
229
230
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using :class:`transformers.XxxTokenizer`.
            See :func:`transformers.PreTrainedTokenizer.encode` and
231
232
233
234
            :func:`transformers.PreTrainedTokenizer.__call__` for details.

            `What are input IDs? <../glossary.html#input-ids>`__
        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
thomwolf's avatar
thomwolf committed
235
236
237
            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.
238
239
240

            `What are attention masks? <../glossary.html#attention-mask>`__
        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
thomwolf's avatar
thomwolf committed
241
242
243
            Segment token indices to indicate first and second portions of the inputs.
            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
            corresponds to a `sentence B` token
244
245
246

            `What are token type IDs? <../glossary.html#token-type-ids>`_
        position_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
thomwolf's avatar
thomwolf committed
247
248
            Indices of positions of each input sequence tokens in the position embeddings.
            Selected in the range ``[0, config.max_position_embeddings - 1]``.
249
250
251

            `What are position IDs? <../glossary.html#position-ids>`_
        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
thomwolf's avatar
thomwolf committed
252
253
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
254
255
256
            :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
257
258
            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.
259
260
261
262
        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.
263
264
265
        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.
thomwolf's avatar
thomwolf committed
266
267
"""

268
269

@add_start_docstrings(
270
    "The bare XXX Model transformer outputting raw hidden-states without any specific head on top.",
271
272
    XXX_START_DOCSTRING,
)
thomwolf's avatar
thomwolf committed
273
274
class XxxModel(XxxPreTrainedModel):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
275
        super().__init__(config)
thomwolf's avatar
thomwolf committed
276
277
278
279
280
281
282

        self.embeddings = XxxEmbeddings(config)
        self.encoder = XxxEncoder(config)
        self.pooler = XxxPooler(config)

        self.init_weights()

thomwolf's avatar
thomwolf committed
283
    def get_input_embeddings(self):
thomwolf's avatar
thomwolf committed
284
285
        return self.embeddings.word_embeddings

thomwolf's avatar
thomwolf committed
286
    def set_input_embeddings(self, new_embeddings):
287
288
        self.embeddings.word_embeddings = new_embeddings

thomwolf's avatar
thomwolf committed
289
290
291
292
293
294
295
296
    def _prune_heads(self, heads_to_prune):
        """ Prunes heads of the model.
            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
            See base class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

297
298
299
300
301
302
303
    @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-uncased",
        output_type=BaseModelOutputWithPooling,
        config_class=_CONFIG_FOR_DOC,
    )
304
305
306
307
308
309
310
311
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
312
313
        output_attentions=None,
        output_hidden_states=None,
314
        return_dict=None,
315
    ):
316
317
318
319
        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
        )
320
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
321

Julien Chaumond's avatar
Julien Chaumond committed
322
323
324
325
326
327
328
329
330
        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 = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

Julien Chaumond's avatar
Julien Chaumond committed
331
332
        device = input_ids.device if input_ids is not None else inputs_embeds.device

thomwolf's avatar
thomwolf committed
333
        if attention_mask is None:
Julien Chaumond's avatar
Julien Chaumond committed
334
            attention_mask = torch.ones(input_shape, device=device)
thomwolf's avatar
thomwolf committed
335
        if token_type_ids is None:
Julien Chaumond's avatar
Julien Chaumond committed
336
            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
thomwolf's avatar
thomwolf committed
337

338
        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
thomwolf's avatar
thomwolf committed
339
340
341
342
343
        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
344
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
thomwolf's avatar
thomwolf committed
345
346
347

        ##################################
        # Replace this with your model code
348
349
350
        embedding_output = self.embeddings(
            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
        )
thomwolf's avatar
thomwolf committed
351
352
        encoder_outputs = self.encoder(embedding_output, extended_attention_mask, head_mask=head_mask)
        sequence_output = encoder_outputs[0]
353
        pooled_output = self.pooler(sequence_output)
thomwolf's avatar
thomwolf committed
354

355
        if not return_dict:
356
            return (sequence_output, pooled_output) + encoder_outputs[1:]
thomwolf's avatar
thomwolf committed
357

358
359
360
361
362
363
        return BaseModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
        )
thomwolf's avatar
thomwolf committed
364

365

366
367
@add_start_docstrings("""XXX Model with a `language modeling` head on top. """, XXX_START_DOCSTRING)
class XxxForMaskedLM(XxxPreTrainedModel):
thomwolf's avatar
thomwolf committed
368
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
369
        super().__init__(config)
thomwolf's avatar
thomwolf committed
370
371

        self.transformer = XxxModel(config)
372
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
thomwolf's avatar
thomwolf committed
373
374
375

        self.init_weights()

thomwolf's avatar
thomwolf committed
376
    def get_output_embeddings(self):
377
        return self.lm_head
thomwolf's avatar
thomwolf committed
378

379
380
381
382
383
384
385
    @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-uncased",
        output_type=MaskedLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
386
387
388
389
390
391
392
393
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
394
395
396
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
397
        return_dict=None,
398
    ):
399
400
401
402
403
404
405
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            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
            in ``[0, ..., config.vocab_size]``
        """
406
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
407
408
409
410
411
412
413
414

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
415
416
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
417
            return_dict=return_dict,
418
        )
thomwolf's avatar
thomwolf committed
419
420
421
422

        sequence_output = outputs[0]
        prediction_scores = self.cls(sequence_output)

423
424
425
426
427
        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()  # -100 index = padding token
            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))

428
        if not return_dict:
429
430
431
432
433
434
435
436
437
            output = (prediction_scores,) + outputs[2:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

        return MaskedLMOutput(
            loss=masked_lm_loss,
            logits=prediction_scores,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
thomwolf's avatar
thomwolf committed
438
439


440
@add_start_docstrings(
441
    """XXX Model transformer with a sequence classification/regression head on top (a linear layer on top of
thomwolf's avatar
thomwolf committed
442
    the pooled output) e.g. for GLUE tasks. """,
443
444
    XXX_START_DOCSTRING,
)
thomwolf's avatar
thomwolf committed
445
446
class XxxForSequenceClassification(XxxPreTrainedModel):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
447
        super().__init__(config)
thomwolf's avatar
thomwolf committed
448
449
450
451
452
453
454
455
        self.num_labels = config.num_labels

        self.transformer = XxxModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)

        self.init_weights()

456
457
458
459
460
461
462
    @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-uncased",
        output_type=SequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
463
464
465
466
467
468
469
470
471
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
472
473
        output_attentions=None,
        output_hidden_states=None,
474
        return_dict=None,
475
    ):
476
477
478
479
480
481
482
        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).
        """
483
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
484
485
486
487
488
489
490
491

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
492
493
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
494
            return_dict=return_dict,
495
        )
thomwolf's avatar
thomwolf committed
496
497
498
499
500
501

        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)

502
        loss = None
thomwolf's avatar
thomwolf committed
503
504
505
506
507
508
509
510
511
        if labels is not None:
            if self.num_labels == 1:
                #  We are doing regression
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), labels.view(-1))
            else:
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

512
        if not return_dict:
513
514
515
516
517
518
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutput(
            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
        )
thomwolf's avatar
thomwolf committed
519
520


521
@add_start_docstrings(
522
523
    """XXX 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. """,
524
525
    XXX_START_DOCSTRING,
)
526
527
528
class XxxForMultipleChoice(XxxPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
thomwolf's avatar
thomwolf committed
529

530
531
532
        self.transformer = XxxModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, 1)
thomwolf's avatar
thomwolf committed
533

534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
        self.init_weights()

    @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-uncased",
        output_type=MultipleChoiceModelOutput,
        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,
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
554
        return_dict=None,
555
556
557
558
559
560
561
    ):
        r"""
        labels (:obj:`torch.LongTensor` 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-1]`` where `num_choices` is the size of the second dimension
            of the input tensors. (see `input_ids` above)
        """
562
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]

        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
        inputs_embeds = (
            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
            if inputs_embeds is not None
            else None
        )

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
584
            return_dict=return_dict,
585
586
587
588
589
590
591
592
593
594
595
596
597
        )

        pooled_output = outputs[1]

        pooled_output = self.dropout(pooled_output)
        logits = self.classifier(pooled_output)
        reshaped_logits = logits.view(-1, num_choices)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(reshaped_logits, labels)

598
        if not return_dict:
599
600
601
602
603
604
            output = (reshaped_logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return MultipleChoiceModelOutput(
            loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
        )
605

606
607
608
609
610
611
612

@add_start_docstrings(
    """XXX Model with a token classification head on top (a linear layer on top of
    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
    XXX_START_DOCSTRING,
)
class XxxForTokenClassification(XxxPreTrainedModel):
thomwolf's avatar
thomwolf committed
613
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
614
        super().__init__(config)
thomwolf's avatar
thomwolf committed
615
616
617
618
619
620
621
622
        self.num_labels = config.num_labels

        self.transformer = XxxModel(config)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        self.init_weights()

623
624
625
626
627
628
629
    @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-uncased",
        output_type=TokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
630
631
632
633
634
635
636
637
638
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
639
640
        output_attentions=None,
        output_hidden_states=None,
641
        return_dict=None,
642
    ):
643
644
645
646
647
        r"""
        labels (:obj:`torch.LongTensor` 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]``.
        """
648
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
649
650
651
652
653
654
655
656

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
657
658
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
659
            return_dict=return_dict,
660
        )
thomwolf's avatar
thomwolf committed
661
662
663
664
665
666

        sequence_output = outputs[0]

        sequence_output = self.dropout(sequence_output)
        logits = self.classifier(sequence_output)

667
        loss = None
thomwolf's avatar
thomwolf committed
668
669
670
671
672
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            # Only keep active parts of the loss
            if attention_mask is not None:
                active_loss = attention_mask.view(-1) == 1
673
674
675
676
                active_logits = logits.view(-1, self.num_labels)
                active_labels = torch.where(
                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
                )
thomwolf's avatar
thomwolf committed
677
678
679
680
                loss = loss_fct(active_logits, active_labels)
            else:
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

681
        if not return_dict:
682
683
684
685
686
687
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions,
        )
thomwolf's avatar
thomwolf committed
688
689


690
@add_start_docstrings(
691
692
    """XXX 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`). """,
693
694
    XXX_START_DOCSTRING,
)
thomwolf's avatar
thomwolf committed
695
696
class XxxForQuestionAnswering(XxxPreTrainedModel):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
697
        super().__init__(config)
thomwolf's avatar
thomwolf committed
698
699
700
701
702
703
704
        self.num_labels = config.num_labels

        self.transformer = XxxModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        self.init_weights()

705
706
707
708
709
710
711
    @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-uncased",
        output_type=QuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
712
713
714
715
716
717
718
719
720
721
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        start_positions=None,
        end_positions=None,
722
723
        output_attentions=None,
        output_hidden_states=None,
724
        return_dict=None,
725
    ):
726
727
728
729
730
731
732
733
734
735
        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.
        """
736
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
737
738
739
740
741
742
743
744

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
745
746
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
747
            return_dict=return_dict,
748
        )
thomwolf's avatar
thomwolf committed
749
750
751
752
753
754
755
756

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1)
        end_logits = end_logits.squeeze(-1)

757
        total_loss = None
thomwolf's avatar
thomwolf committed
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions.clamp_(0, ignored_index)
            end_positions.clamp_(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

774
        if not return_dict:
775
776
777
778
779
780
781
782
783
784
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )