modeling_xxx.py 31.4 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
####################################################
# 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):
Lysandre's avatar
Lysandre committed
62
    """Load tf checkpoints in a pytorch model."""
thomwolf's avatar
thomwolf committed
63
64
    try:
        import re
65

thomwolf's avatar
thomwolf committed
66
67
68
        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
            array = np.transpose(array)
        try:
Teven's avatar
Teven committed
124
125
126
            assert (
                pointer.shape == array.shape
            ), f"Pointer and array have mismatched shapes {pointer.shape} and {array.shape}"
thomwolf's avatar
thomwolf committed
127
128
129
130
131
132
133
134
135
136
137
        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
138
# - PreTrainedModel for the models (itself a sub-class of torch.nn.Module)
thomwolf's avatar
thomwolf committed
139
140
141
142
143
144
145
146
####################################################

####################################################
# 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
####################################################
147
148
149
150
151
152
153
154

XxxAttention = nn.Module

XxxIntermediate = nn.Module

XxxOutput = nn.Module


thomwolf's avatar
thomwolf committed
155
156
class XxxLayer(nn.Module):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
157
        super().__init__()
thomwolf's avatar
thomwolf committed
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
        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
####################################################
180
181
182
183
184
185
186
187
188
189

XxxLayerNorm = torch.nn.LayerNorm

XxxEmbeddings = nn.Module

XxxEncoder = nn.Module

XxxPooler = nn.Module


thomwolf's avatar
thomwolf committed
190
class XxxPreTrainedModel(PreTrainedModel):
Lysandre's avatar
Lysandre committed
191
192
    """An abstract class to handle weights initialization and
    a simple interface for downloading and loading pretrained models.
thomwolf's avatar
thomwolf committed
193
    """
194

thomwolf's avatar
thomwolf committed
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
    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
213
214
    `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding
    <https://arxiv.org/abs/1810.04805>`__ by....
thomwolf's avatar
thomwolf committed
215

216
217
218
    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
219
220

    Parameters:
221
        config (:class:`~transformers.XxxConfig`): Model configuration class with all the parameters of the model.
thomwolf's avatar
thomwolf committed
222
223
224
225
226
227
            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:
228
        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
thomwolf's avatar
thomwolf committed
229
230
231
232
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using :class:`transformers.XxxTokenizer`.
            See :func:`transformers.PreTrainedTokenizer.encode` and
233
234
235
236
            :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
237
238
239
            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.
240
241
242

            `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
243
244
245
            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
246
247
248

            `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
249
250
            Indices of positions of each input sequence tokens in the position embeddings.
            Selected in the range ``[0, config.max_position_embeddings - 1]``.
251
252
253

            `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
254
255
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
256
257
258
            :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.
259
260
            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.
261
262
263
264
        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.
265
266
267
        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
268
269
"""

270
271

@add_start_docstrings(
272
    "The bare XXX Model transformer outputting raw hidden-states without any specific head on top.",
273
274
    XXX_START_DOCSTRING,
)
thomwolf's avatar
thomwolf committed
275
276
class XxxModel(XxxPreTrainedModel):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
277
        super().__init__(config)
thomwolf's avatar
thomwolf committed
278
279
280
281
282
283
284

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

        self.init_weights()

thomwolf's avatar
thomwolf committed
285
    def get_input_embeddings(self):
thomwolf's avatar
thomwolf committed
286
287
        return self.embeddings.word_embeddings

thomwolf's avatar
thomwolf committed
288
    def set_input_embeddings(self, new_embeddings):
289
290
        self.embeddings.word_embeddings = new_embeddings

thomwolf's avatar
thomwolf committed
291
    def _prune_heads(self, heads_to_prune):
Lysandre's avatar
Lysandre committed
292
293
294
        """Prunes heads of the model.
        heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
        See base class PreTrainedModel
thomwolf's avatar
thomwolf committed
295
296
297
298
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

299
300
301
302
303
304
305
    @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,
    )
306
307
308
309
310
311
312
313
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
314
315
        output_attentions=None,
        output_hidden_states=None,
316
        return_dict=None,
317
    ):
318
319
320
321
        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
        )
322
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
323

Julien Chaumond's avatar
Julien Chaumond committed
324
325
326
327
328
329
330
331
332
        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
333
334
        device = input_ids.device if input_ids is not None else inputs_embeds.device

thomwolf's avatar
thomwolf committed
335
        if attention_mask is None:
Julien Chaumond's avatar
Julien Chaumond committed
336
            attention_mask = torch.ones(input_shape, device=device)
thomwolf's avatar
thomwolf committed
337
        if token_type_ids is None:
Julien Chaumond's avatar
Julien Chaumond committed
338
            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
thomwolf's avatar
thomwolf committed
339

340
        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
thomwolf's avatar
thomwolf committed
341
342
343
344
345
        # 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]
346
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
thomwolf's avatar
thomwolf committed
347
348
349

        ##################################
        # Replace this with your model code
350
351
352
        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
353
354
        encoder_outputs = self.encoder(embedding_output, extended_attention_mask, head_mask=head_mask)
        sequence_output = encoder_outputs[0]
355
        pooled_output = self.pooler(sequence_output)
thomwolf's avatar
thomwolf committed
356

357
        if not return_dict:
358
            return (sequence_output, pooled_output) + encoder_outputs[1:]
thomwolf's avatar
thomwolf committed
359

360
361
362
363
364
365
        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
366

367

368
369
@add_start_docstrings("""XXX Model with a `language modeling` head on top. """, XXX_START_DOCSTRING)
class XxxForMaskedLM(XxxPreTrainedModel):
thomwolf's avatar
thomwolf committed
370
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
371
        super().__init__(config)
thomwolf's avatar
thomwolf committed
372
373

        self.transformer = XxxModel(config)
374
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
thomwolf's avatar
thomwolf committed
375
376
377

        self.init_weights()

thomwolf's avatar
thomwolf committed
378
    def get_output_embeddings(self):
379
        return self.lm_head
thomwolf's avatar
thomwolf committed
380

381
382
383
384
385
386
387
    @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,
    )
388
389
390
391
392
393
394
395
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
396
397
398
        labels=None,
        output_attentions=None,
        output_hidden_states=None,
399
        return_dict=None,
400
    ):
401
402
403
404
405
406
407
        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]``
        """
408
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
409
410
411
412
413
414
415
416

        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,
417
418
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
419
            return_dict=return_dict,
420
        )
thomwolf's avatar
thomwolf committed
421
422
423
424

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

425
426
427
428
429
        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))

430
        if not return_dict:
431
432
433
434
435
436
437
438
439
            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
440
441


442
@add_start_docstrings(
443
    """XXX Model transformer with a sequence classification/regression head on top (a linear layer on top of
thomwolf's avatar
thomwolf committed
444
    the pooled output) e.g. for GLUE tasks. """,
445
446
    XXX_START_DOCSTRING,
)
thomwolf's avatar
thomwolf committed
447
448
class XxxForSequenceClassification(XxxPreTrainedModel):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
449
        super().__init__(config)
thomwolf's avatar
thomwolf committed
450
451
452
453
454
455
456
457
        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()

458
459
460
461
462
463
464
    @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,
    )
465
466
467
468
469
470
471
472
473
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
474
475
        output_attentions=None,
        output_hidden_states=None,
476
        return_dict=None,
477
    ):
478
479
480
481
482
483
484
        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).
        """
485
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
486
487
488
489
490
491
492
493

        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,
494
495
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
496
            return_dict=return_dict,
497
        )
thomwolf's avatar
thomwolf committed
498
499
500
501
502
503

        pooled_output = outputs[1]

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

504
        loss = None
thomwolf's avatar
thomwolf committed
505
506
507
508
509
510
511
512
513
        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))

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

        return SequenceClassifierOutput(
Lysandre's avatar
Lysandre committed
519
520
521
522
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
523
        )
thomwolf's avatar
thomwolf committed
524
525


526
@add_start_docstrings(
527
528
    """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. """,
529
530
    XXX_START_DOCSTRING,
)
531
532
533
class XxxForMultipleChoice(XxxPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
thomwolf's avatar
thomwolf committed
534

535
536
537
        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
538

539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
        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,
559
        return_dict=None,
560
561
562
563
564
565
566
    ):
        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)
        """
567
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
        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,
589
            return_dict=return_dict,
590
591
592
593
594
595
596
597
598
599
600
601
602
        )

        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)

603
        if not return_dict:
604
605
606
607
            output = (reshaped_logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return MultipleChoiceModelOutput(
Lysandre's avatar
Lysandre committed
608
609
610
611
            loss=loss,
            logits=reshaped_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
612
        )
613

614
615
616
617
618
619
620

@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
621
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
622
        super().__init__(config)
thomwolf's avatar
thomwolf committed
623
624
625
626
627
628
629
630
        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()

631
632
633
634
635
636
637
    @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,
    )
638
639
640
641
642
643
644
645
646
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
647
648
        output_attentions=None,
        output_hidden_states=None,
649
        return_dict=None,
650
    ):
651
652
653
654
655
        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]``.
        """
656
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
657
658
659
660
661
662
663
664

        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,
665
666
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
667
            return_dict=return_dict,
668
        )
thomwolf's avatar
thomwolf committed
669
670
671
672
673
674

        sequence_output = outputs[0]

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

675
        loss = None
thomwolf's avatar
thomwolf committed
676
677
678
679
680
        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
681
682
683
684
                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
685
686
687
688
                loss = loss_fct(active_logits, active_labels)
            else:
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

689
        if not return_dict:
690
691
692
693
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
Lysandre's avatar
Lysandre committed
694
695
696
697
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
698
        )
thomwolf's avatar
thomwolf committed
699
700


701
@add_start_docstrings(
702
703
    """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`). """,
704
705
    XXX_START_DOCSTRING,
)
thomwolf's avatar
thomwolf committed
706
707
class XxxForQuestionAnswering(XxxPreTrainedModel):
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
708
        super().__init__(config)
thomwolf's avatar
thomwolf committed
709
710
711
712
713
714
715
        self.num_labels = config.num_labels

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

        self.init_weights()

716
717
718
719
720
721
722
    @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,
    )
723
724
725
726
727
728
729
730
731
732
    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,
733
734
        output_attentions=None,
        output_hidden_states=None,
735
        return_dict=None,
736
    ):
737
738
739
740
741
742
743
744
745
746
        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.
        """
747
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
748
749
750
751
752
753
754
755

        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,
756
757
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
758
            return_dict=return_dict,
759
        )
thomwolf's avatar
thomwolf committed
760
761
762
763
764
765
766
767

        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)

768
        total_loss = None
thomwolf's avatar
thomwolf committed
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
        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

785
        if not return_dict:
786
787
788
789
790
791
792
793
794
795
            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,
        )