modeling_utils.py 47.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# 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 BERT model."""

18
19
from __future__ import (absolute_import, division, print_function,
                        unicode_literals)
20

21
22
import copy
import json
23
24
import logging
import os
thomwolf's avatar
thomwolf committed
25
from io import open
26

27
import six
28
29
import torch
from torch import nn
30
31
from torch.nn import CrossEntropyLoss
from torch.nn import functional as F
32
33
34
35
36
37
38

from .file_utils import cached_path

logger = logging.getLogger(__name__)

CONFIG_NAME = "config.json"
WEIGHTS_NAME = "pytorch_model.bin"
39
TF_WEIGHTS_NAME = 'model.ckpt'
40
41


thomwolf's avatar
thomwolf committed
42
43
44
45
46
47
48
49
50
51
52
53
54
55
try:
    from torch.nn import Identity
except ImportError:
    # Older PyTorch compatibility
    class Identity(nn.Module):
        r"""A placeholder identity operator that is argument-insensitive.
        """
        def __init__(self, *args, **kwargs):
            super(Identity, self).__init__()

        def forward(self, input):
            return input


56
57
58
59
60
61
62
63
64
65
66
67
if not six.PY2:
    def add_start_docstrings(*docstr):
        def docstring_decorator(fn):
            fn.__doc__ = ''.join(docstr) + fn.__doc__
            return fn
        return docstring_decorator
else:
    # Not possible to update class docstrings on python2
    def add_start_docstrings(*docstr):
        def docstring_decorator(fn):
            return fn
        return docstring_decorator
thomwolf's avatar
thomwolf committed
68
69


70
class PretrainedConfig(object):
71
72
73
    r""" Base class for all configuration classes.
        Handles a few parameters common to all models' configurations as well as methods for loading/downloading/saving configurations.

74
75
76
77
        Note:
            A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to initialize a model does **not** load the model weights.
            It only affects the model's configuration.

78
79
80
81
82
83
84
85
86
        Class attributes (overridden by derived classes):
            - ``pretrained_config_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained model configurations as values.

        Parameters:
            ``finetuning_task``: string, default `None`. Name of the task used to fine-tune the model. This can be used when converting from an original (TensorFlow or PyTorch) checkpoint.
            ``num_labels``: integer, default `2`. Number of classes to use when the model is a classification model (sequences/tokens)
            ``output_attentions``: boolean, default `False`. Should the model returns attentions weights.
            ``output_hidden_states``: string, default `False`. Should the model returns all hidden-states.
            ``torchscript``: string, default `False`. Is the model used with Torchscript.
87
88
89
    """
    pretrained_config_archive_map = {}

thomwolf's avatar
thomwolf committed
90
91
92
93
94
    def __init__(self, **kwargs):
        self.finetuning_task = kwargs.pop('finetuning_task', None)
        self.num_labels = kwargs.pop('num_labels', 2)
        self.output_attentions = kwargs.pop('output_attentions', False)
        self.output_hidden_states = kwargs.pop('output_hidden_states', False)
95
        self.torchscript = kwargs.pop('torchscript', False)
thomwolf's avatar
thomwolf committed
96

thomwolf's avatar
thomwolf committed
97
    def save_pretrained(self, save_directory):
thomwolf's avatar
thomwolf committed
98
        """ Save a configuration object to the directory `save_directory`, so that it
99
            can be re-loaded using the :func:`~pytorch_transformers.PretrainedConfig.from_pretrained` class method.
thomwolf's avatar
thomwolf committed
100
101
102
103
104
105
106
107
        """
        assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"

        # If we save using the predefined names, we can load using `from_pretrained`
        output_config_file = os.path.join(save_directory, CONFIG_NAME)

        self.to_json_file(output_config_file)

108
    @classmethod
109
    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
110
        r""" Instantiate a :class:`~pytorch_transformers.PretrainedConfig` (or a derived class) from a pre-trained model configuration.
111

thomwolf's avatar
thomwolf committed
112
        Parameters:
113
            pretrained_model_name_or_path: either:
thomwolf's avatar
thomwolf committed
114
115

                - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
116
                - a path to a `directory` containing a configuration file saved using the :func:`~pytorch_transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
thomwolf's avatar
thomwolf committed
117
118
                - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.

119
            cache_dir: (`optional`) string:
thomwolf's avatar
thomwolf committed
120
121
                Path to a directory in which a downloaded pre-trained model
                configuration should be cached if the standard cache should not be used.
thomwolf's avatar
thomwolf committed
122

123
            kwargs: (`optional`) dict: key/value pairs with which to update the configuration object after loading.
thomwolf's avatar
thomwolf committed
124

125
126
                - The values in kwargs of any keys which are configuration attributes will be used to override the loaded values.
                - Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled by the `return_unused_kwargs` keyword parameter.
thomwolf's avatar
thomwolf committed
127

128
129
130
            force_download: (`optional`) boolean, default False:
                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.

131
            return_unused_kwargs: (`optional`) bool:
thomwolf's avatar
thomwolf committed
132

133
134
                - If False, then this function returns just the final configuration object.
                - If True, then this functions returns a tuple `(config, unused_kwargs)` where `unused_kwargs` is a dictionary consisting of the key/value pairs whose keys are not configuration attributes: ie the part of kwargs which has not been used to update `config` and is otherwise ignored.
thomwolf's avatar
thomwolf committed
135
136
137

        Examples::

thomwolf's avatar
thomwolf committed
138
139
            # We can't instantiate directly the base class `PretrainedConfig` so let's show the examples on a
            # derived class: BertConfig
thomwolf's avatar
thomwolf committed
140
141
142
143
144
145
146
147
148
            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
            config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
            config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
            config = BertConfig.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
            assert config.output_attention == True
            config, unused_kwargs = BertConfig.from_pretrained('bert-base-uncased', output_attention=True,
                                                               foo=False, return_unused_kwargs=True)
            assert config.output_attention == True
            assert unused_kwargs == {'foo': False}
thomwolf's avatar
thomwolf committed
149

150
        """
151
        cache_dir = kwargs.pop('cache_dir', None)
152
        force_download = kwargs.pop('force_download', False)
153
        return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
154
155
156

        if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
            config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
thomwolf's avatar
thomwolf committed
157
        elif os.path.isdir(pretrained_model_name_or_path):
158
            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
thomwolf's avatar
thomwolf committed
159
160
        else:
            config_file = pretrained_model_name_or_path
161
162
        # redirect to the cache, if necessary
        try:
163
            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download)
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
        except EnvironmentError:
            if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
                        config_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find any file "
                    "associated to this path or url.".format(
                        pretrained_model_name_or_path,
                        ', '.join(cls.pretrained_config_archive_map.keys()),
                        config_file))
            return None
        if resolved_config_file == config_file:
            logger.info("loading configuration file {}".format(config_file))
        else:
            logger.info("loading configuration file {} from cache at {}".format(
                config_file, resolved_config_file))

        # Load config
        config = cls.from_json_file(resolved_config_file)

        # Update config with kwargs if needed
        to_remove = []
        for key, value in kwargs.items():
            if hasattr(config, key):
                setattr(config, key, value)
                to_remove.append(key)
        for key in to_remove:
            kwargs.pop(key, None)

thomwolf's avatar
thomwolf committed
196
        logger.info("Model config %s", config)
197
        if return_unused_kwargs:
198
199
200
            return config, kwargs
        else:
            return config
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216

    @classmethod
    def from_dict(cls, json_object):
        """Constructs a `Config` from a Python dictionary of parameters."""
        config = cls(vocab_size_or_config_json_file=-1)
        for key, value in json_object.items():
            config.__dict__[key] = value
        return config

    @classmethod
    def from_json_file(cls, json_file):
        """Constructs a `BertConfig` from a json file of parameters."""
        with open(json_file, "r", encoding='utf-8') as reader:
            text = reader.read()
        return cls.from_dict(json.loads(text))

thomwolf's avatar
thomwolf committed
217
218
219
    def __eq__(self, other):
        return self.__dict__ == other.__dict__

220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
    def __repr__(self):
        return str(self.to_json_string())

    def to_dict(self):
        """Serializes this instance to a Python dictionary."""
        output = copy.deepcopy(self.__dict__)
        return output

    def to_json_string(self):
        """Serializes this instance to a JSON string."""
        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"

    def to_json_file(self, json_file_path):
        """ Save this instance to a json file."""
        with open(json_file_path, "w", encoding='utf-8') as writer:
            writer.write(self.to_json_string())


238
class PreTrainedModel(nn.Module):
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
    r""" Base class for all models.

        :class:`~pytorch_transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
        as well as a few methods commons to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.

        Class attributes (overridden by derived classes):
            - ``config_class``: a class derived from :class:`~pytorch_transformers.PretrainedConfig` to use as configuration class for this model architecture.
            - ``pretrained_model_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained weights as values.
            - ``load_tf_weights``: a python ``method`` for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments:

                - ``model``: an instance of the relevant subclass of :class:`~pytorch_transformers.PreTrainedModel`,
                - ``config``: an instance of the relevant subclass of :class:`~pytorch_transformers.PretrainedConfig`,
                - ``path``: a path (string) to the TensorFlow checkpoint.

            - ``base_model_prefix``: a string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.
254
    """
255
    config_class = None
256
257
258
259
260
261
262
263
264
265
266
267
268
    pretrained_model_archive_map = {}
    load_tf_weights = lambda model, config, path: None
    base_model_prefix = ""

    def __init__(self, config, *inputs, **kwargs):
        super(PreTrainedModel, self).__init__()
        if not isinstance(config, PretrainedConfig):
            raise ValueError(
                "Parameter config in `{}(config)` should be an instance of class `PretrainedConfig`. "
                "To create a model from a pretrained model use "
                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
                    self.__class__.__name__, self.__class__.__name__
                ))
thomwolf's avatar
thomwolf committed
269
        # Save config in model
270
271
        self.config = config

thomwolf's avatar
thomwolf committed
272
273
274
275
276
277
    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
        """ Build a resized Embedding Module from a provided token Embedding Module.
            Increasing the size will add newly initialized vectors at the end
            Reducing the size will remove vectors from the end

        Args:
thomwolf's avatar
thomwolf committed
278
279
            new_num_tokens: (`optional`) int
                New number of tokens in the embedding matrix.
thomwolf's avatar
thomwolf committed
280
281
282
                Increasing the size will add newly initialized vectors at the end
                Reducing the size will remove vectors from the end
                If not provided or None: return the provided token Embedding Module.
thomwolf's avatar
thomwolf committed
283
        Return: ``torch.nn.Embeddings``
thomwolf's avatar
thomwolf committed
284
285
286
287
288
            Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None
        """
        if new_num_tokens is None:
            return old_embeddings

thomwolf's avatar
thomwolf committed
289
        old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
thomwolf's avatar
thomwolf committed
290
291
292
293
        if old_num_tokens == new_num_tokens:
            return old_embeddings

        # Build new embeddings
thomwolf's avatar
thomwolf committed
294
295
296
297
298
299
300
301
302
303
304
305
        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
        new_embeddings.to(old_embeddings.weight.device)

        # initialize all new embeddings (in particular added tokens)
        self.init_weights(new_embeddings)

        # Copy word embeddings from the previous weights
        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
        new_embeddings.weight.data[:num_tokens_to_copy, :] = old_embeddings.weight.data[:num_tokens_to_copy, :]

        return new_embeddings

thomwolf's avatar
thomwolf committed
306
307
308
309
310
311
312
313
314
315
    def _tie_or_clone_weights(self, first_module, second_module):
        """ Tie or clone module weights depending of weither we are using TorchScript or not
        """
        if self.config.torchscript:
            first_module.weight = nn.Parameter(second_module.weight.clone())
        else:
            first_module.weight = second_module.weight

    def resize_token_embeddings(self, new_num_tokens=None):
        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
316
        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
thomwolf's avatar
thomwolf committed
317

318
319
320
321
322
        Arguments:

            new_num_tokens: (`optional`) int:
                New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end. 
                If not provided or None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.
thomwolf's avatar
thomwolf committed
323

thomwolf's avatar
thomwolf committed
324
        Return: ``torch.nn.Embeddings``
325
            Pointer to the input tokens Embeddings Module of the model
thomwolf's avatar
thomwolf committed
326
327
        """
        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
thomwolf's avatar
thomwolf committed
328
329
330
        model_embeds = base_model._resize_token_embeddings(new_num_tokens)
        if new_num_tokens is None:
            return model_embeds
thomwolf's avatar
thomwolf committed
331
332
333
334
335
336
337
338
339

        # Update base model and current model config
        self.config.vocab_size = new_num_tokens
        base_model.vocab_size = new_num_tokens

        # Tie weights again if needed
        if hasattr(self, 'tie_weights'):
            self.tie_weights()

thomwolf's avatar
thomwolf committed
340
341
        return model_embeds

thomwolf's avatar
thomwolf committed
342
343
    def prune_heads(self, heads_to_prune):
        """ Prunes heads of the base model.
344
345
346
347

            Arguments:

                heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
thomwolf's avatar
thomwolf committed
348
        """
thomwolf's avatar
thomwolf committed
349
350
        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
        base_model._prune_heads(heads_to_prune)
thomwolf's avatar
thomwolf committed
351

352
    def save_pretrained(self, save_directory):
353
354
        """ Save a model and its configuration file to a directory, so that it
            can be re-loaded using the `:func:`~pytorch_transformers.PreTrainedModel.from_pretrained`` class method.
355
356
357
358
359
360
        """
        assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"

        # Only save the model it-self if we are using distributed training
        model_to_save = self.module if hasattr(self, 'module') else self

thomwolf's avatar
thomwolf committed
361
362
363
        # Save configuration file
        model_to_save.config.save_pretrained(save_directory)

364
365
366
367
368
        # If we save using the predefined names, we can load using `from_pretrained`
        output_model_file = os.path.join(save_directory, WEIGHTS_NAME)

        torch.save(model_to_save.state_dict(), output_model_file)

369
    @classmethod
370
    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
371
372
        r"""Instantiate a pretrained pytorch model from a pre-trained model configuration.

373
374
375
        The model is set in evaluation mode by default using ``model.eval()`` (Dropout modules are deactivated)
        To train the model, you should first set it back in training mode with ``model.train()``

376
377
378
379
380
        The warning ``Weights from XXX not initialized from pretrained model`` means that the weights of XXX do not come pre-trained with the rest of the model.
        It is up to you to train those weights with a downstream fine-tuning task.

        The warning ``Weights from XXX not used in YYY`` means that the layer XXX is not used by YYY, therefore those weights are discarded.

381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
        Parameters:
            pretrained_model_name_or_path: either:

                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
                - a path to a `directory` containing model weights saved using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.

            model_args: (`optional`) Sequence of positional arguments:
                All remaning positional arguments will be passed to the underlying model's ``__init__`` method

            config: (`optional`) instance of a class derived from :class:`~pytorch_transformers.PretrainedConfig`:
                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:

                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
                - the model was saved using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.

            state_dict: (`optional`) dict:
                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
thomwolf's avatar
typos  
thomwolf committed
400
                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
401
402
403
                In this case though, you should check if using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained` and :func:`~pytorch_transformers.PreTrainedModel.from_pretrained` is not a simpler option.

            cache_dir: (`optional`) string:
thomwolf's avatar
thomwolf committed
404
405
                Path to a directory in which a downloaded pre-trained model
                configuration should be cached if the standard cache should not be used.
406

407
408
409
            force_download: (`optional`) boolean, default False:
                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.

410
            output_loading_info: (`optional`) boolean:
thomwolf's avatar
thomwolf committed
411
                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
412
413
414
415
416
417
418
419

            kwargs: (`optional`) Remaining dictionary of keyword arguments:
                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:

                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~pytorch_transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.

        Examples::
thomwolf's avatar
thomwolf committed
420

thomwolf's avatar
thomwolf committed
421
422
423
424
425
426
427
            model = BertModel.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
            model = BertModel.from_pretrained('./test/saved_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
            model = BertModel.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
            assert model.config.output_attention == True
            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
            config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
            model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
thomwolf's avatar
thomwolf committed
428

429
        """
thomwolf's avatar
thomwolf committed
430
        config = kwargs.pop('config', None)
thomwolf's avatar
thomwolf committed
431
432
        state_dict = kwargs.pop('state_dict', None)
        cache_dir = kwargs.pop('cache_dir', None)
thomwolf's avatar
thomwolf committed
433
        from_tf = kwargs.pop('from_tf', False)
434
        force_download = kwargs.pop('force_download', False)
thomwolf's avatar
thomwolf committed
435
        output_loading_info = kwargs.pop('output_loading_info', False)
thomwolf's avatar
thomwolf committed
436
437

        # Load config
thomwolf's avatar
thomwolf committed
438
        if config is None:
439
440
            config, model_kwargs = cls.config_class.from_pretrained(
                pretrained_model_name_or_path, *model_args,
441
                cache_dir=cache_dir, return_unused_kwargs=True,
442
                force_download=force_download,
443
                **kwargs
444
445
446
            )
        else:
            model_kwargs = kwargs
447

thomwolf's avatar
thomwolf committed
448
        # Load model
449
450
        if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
            archive_file = cls.pretrained_model_archive_map[pretrained_model_name_or_path]
thomwolf's avatar
thomwolf committed
451
        elif os.path.isdir(pretrained_model_name_or_path):
452
453
454
455
456
            if from_tf:
                # Directly load from a TensorFlow checkpoint
                archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
            else:
                archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
thomwolf's avatar
thomwolf committed
457
458
459
460
461
462
        else:
            if from_tf:
                # Directly load from a TensorFlow checkpoint
                archive_file = pretrained_model_name_or_path + ".index"
            else:
                archive_file = pretrained_model_name_or_path
463
464
        # redirect to the cache, if necessary
        try:
465
            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download)
466
467
468
469
470
471
472
473
474
475
476
477
478
479
        except EnvironmentError:
            if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained weights.".format(
                        archive_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find any file "
                    "associated to this path or url.".format(
                        pretrained_model_name_or_path,
                        ', '.join(cls.pretrained_model_archive_map.keys()),
                        archive_file))
            return None
thomwolf's avatar
thomwolf committed
480
        if resolved_archive_file == archive_file:
481
482
483
484
485
486
            logger.info("loading weights file {}".format(archive_file))
        else:
            logger.info("loading weights file {} from cache at {}".format(
                archive_file, resolved_archive_file))

        # Instantiate model.
487
        model = cls(config, *model_args, **model_kwargs)
thomwolf's avatar
thomwolf committed
488

489
490
491
492
        if state_dict is None and not from_tf:
            state_dict = torch.load(resolved_archive_file, map_location='cpu')
        if from_tf:
            # Directly load from a TensorFlow checkpoint
thomwolf's avatar
thomwolf committed
493
            return cls.load_tf_weights(model, config, resolved_archive_file[:-6])  # Remove the '.index'
494

thomwolf's avatar
thomwolf committed
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
        # Convert old format to new format if needed from a PyTorch state_dict
        old_keys = []
        new_keys = []
        for key in state_dict.keys():
            new_key = None
            if 'gamma' in key:
                new_key = key.replace('gamma', 'weight')
            if 'beta' in key:
                new_key = key.replace('beta', 'bias')
            if new_key:
                old_keys.append(key)
                new_keys.append(new_key)
        for old_key, new_key in zip(old_keys, new_keys):
            state_dict[new_key] = state_dict.pop(old_key)

510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
        # Load from a PyTorch state_dict
        missing_keys = []
        unexpected_keys = []
        error_msgs = []
        # copy state_dict so _load_from_state_dict can modify it
        metadata = getattr(state_dict, '_metadata', None)
        state_dict = state_dict.copy()
        if metadata is not None:
            state_dict._metadata = metadata

        def load(module, prefix=''):
            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
            module._load_from_state_dict(
                state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)
            for name, child in module._modules.items():
                if child is not None:
                    load(child, prefix + name + '.')

thomwolf's avatar
thomwolf committed
528
        # Make sure we are able to load base models as well as derived models (with heads)
529
        start_prefix = ''
thomwolf's avatar
thomwolf committed
530
        model_to_load = model
531
        if not hasattr(model, cls.base_model_prefix) and any(s.startswith(cls.base_model_prefix) for s in state_dict.keys()):
thomwolf's avatar
thomwolf committed
532
533
534
535
536
            start_prefix = cls.base_model_prefix + '.'
        if hasattr(model, cls.base_model_prefix) and not any(s.startswith(cls.base_model_prefix) for s in state_dict.keys()):
            model_to_load = getattr(model, cls.base_model_prefix)

        load(model_to_load, prefix=start_prefix)
537
538
539
540
541
542
543
544
545
546
        if len(missing_keys) > 0:
            logger.info("Weights of {} not initialized from pretrained model: {}".format(
                model.__class__.__name__, missing_keys))
        if len(unexpected_keys) > 0:
            logger.info("Weights from pretrained model not used in {}: {}".format(
                model.__class__.__name__, unexpected_keys))
        if len(error_msgs) > 0:
            raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
                               model.__class__.__name__, "\n\t".join(error_msgs)))

thomwolf's avatar
thomwolf committed
547
        if hasattr(model, 'tie_weights'):
548
549
            model.tie_weights()  # make sure word embedding weights are still tied

550
551
552
        # Set model in evaluation mode to desactivate DropOut modules by default
        model.eval()

thomwolf's avatar
thomwolf committed
553
554
555
556
        if output_loading_info:
            loading_info = {"missing_keys": missing_keys, "unexpected_keys": unexpected_keys, "error_msgs": error_msgs}
            return model, loading_info

557
558
559
        return model


thomwolf's avatar
thomwolf committed
560
561
class Conv1D(nn.Module):
    def __init__(self, nf, nx):
thomwolf's avatar
thomwolf committed
562
        """ Conv1D layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2)
thomwolf's avatar
thomwolf committed
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
            Basically works like a Linear layer but the weights are transposed
        """
        super(Conv1D, self).__init__()
        self.nf = nf
        w = torch.empty(nx, nf)
        nn.init.normal_(w, std=0.02)
        self.weight = nn.Parameter(w)
        self.bias = nn.Parameter(torch.zeros(nf))

    def forward(self, x):
        size_out = x.size()[:-1] + (self.nf,)
        x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
        x = x.view(*size_out)
        return x


thomwolf's avatar
thomwolf committed
579
580
class PoolerStartLogits(nn.Module):
    """ Compute SQuAD start_logits from sequence hidden states. """
thomwolf's avatar
thomwolf committed
581
    def __init__(self, config):
thomwolf's avatar
thomwolf committed
582
583
584
585
586
        super(PoolerStartLogits, self).__init__()
        self.dense = nn.Linear(config.hidden_size, 1)

    def forward(self, hidden_states, p_mask=None):
        """ Args:
587
588
589
            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape `(batch_size, seq_len)`
                invalid position mask such as query and special symbols (PAD, SEP, CLS)
                1.0 means token should be masked.
thomwolf's avatar
thomwolf committed
590
        """
thomwolf's avatar
thomwolf committed
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
            x = x * (1 - p_mask) - 1e30 * p_mask

        return x


class PoolerEndLogits(nn.Module):
    """ Compute SQuAD end_logits from sequence hidden states and start token hidden state.
    """
    def __init__(self, config):
        super(PoolerEndLogits, self).__init__()
        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
        self.activation = nn.Tanh()
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dense_1 = nn.Linear(config.hidden_size, 1)

    def forward(self, hidden_states, start_states=None, start_positions=None, p_mask=None):
        """ Args:
611
612
613
614
615
616
            One of ``start_states``, ``start_positions`` should be not None.
            If both are set, ``start_positions`` overrides ``start_states``.

            **start_states**: ``torch.LongTensor`` of shape identical to hidden_states
                hidden states of the first tokens for the labeled span.
            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
617
                position of the first token for the labeled span:
618
619
620
            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
                Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
                1.0 means token should be masked.
thomwolf's avatar
thomwolf committed
621
622
623
        """
        assert start_states is not None or start_positions is not None, "One of start_states, start_positions should be not None"
        if start_positions is not None:
624
            slen, hsz = hidden_states.shape[-2:]
thomwolf's avatar
thomwolf committed
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
            start_positions = start_positions[:, None, None].expand(-1, -1, hsz) # shape (bsz, 1, hsz)
            start_states = hidden_states.gather(-2, start_positions) # shape (bsz, 1, hsz)
            start_states = start_states.expand(-1, slen, -1) # shape (bsz, slen, hsz)

        x = self.dense_0(torch.cat([hidden_states, start_states], dim=-1))
        x = self.activation(x)
        x = self.LayerNorm(x)
        x = self.dense_1(x).squeeze(-1)

        if p_mask is not None:
            x = x * (1 - p_mask) - 1e30 * p_mask

        return x


class PoolerAnswerClass(nn.Module):
    """ Compute SQuAD 2.0 answer class from classification and start tokens hidden states. """
    def __init__(self, config):
        super(PoolerAnswerClass, self).__init__()
        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
        self.activation = nn.Tanh()
        self.dense_1 = nn.Linear(config.hidden_size, 1, bias=False)

    def forward(self, hidden_states, start_states=None, start_positions=None, cls_index=None):
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
        """
        Args:
            One of ``start_states``, ``start_positions`` should be not None.
            If both are set, ``start_positions`` overrides ``start_states``.

            **start_states**: ``torch.LongTensor`` of shape identical to ``hidden_states``.
                hidden states of the first tokens for the labeled span.
            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
                position of the first token for the labeled span.
            **cls_index**: torch.LongTensor of shape ``(batch_size,)``
                position of the CLS token. If None, take the last token.

            note(Original repo):
                no dependency on end_feature so that we can obtain one single `cls_logits`
                for each sample
thomwolf's avatar
thomwolf committed
664
        """
665
        hsz = hidden_states.shape[-1]
thomwolf's avatar
thomwolf committed
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
        assert start_states is not None or start_positions is not None, "One of start_states, start_positions should be not None"
        if start_positions is not None:
            start_positions = start_positions[:, None, None].expand(-1, -1, hsz) # shape (bsz, 1, hsz)
            start_states = hidden_states.gather(-2, start_positions).squeeze(-2) # shape (bsz, hsz)

        if cls_index is not None:
            cls_index = cls_index[:, None, None].expand(-1, -1, hsz) # shape (bsz, 1, hsz)
            cls_token_state = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, hsz)
        else:
            cls_token_state = hidden_states[:, -1, :] # shape (bsz, hsz)

        x = self.dense_0(torch.cat([start_states, cls_token_state], dim=-1))
        x = self.activation(x)
        x = self.dense_1(x).squeeze(-1)

        return x


class SQuADHead(nn.Module):
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
    r""" A SQuAD head inspired by XLNet.

    Parameters:
        config (:class:`~pytorch_transformers.XLNetConfig`): Model configuration class with all the parameters of the model.

    Inputs:
        **hidden_states**: ``torch.FloatTensor`` of shape ``(batch_size, seq_len, hidden_size)``
            hidden states of sequence tokens
        **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
            position of the first token for the labeled span.
        **end_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
            position of the last token for the labeled span.
        **cls_index**: torch.LongTensor of shape ``(batch_size,)``
            position of the CLS token. If None, take the last token.
        **is_impossible**: ``torch.LongTensor`` of shape ``(batch_size,)``
            Whether the question has a possible answer in the paragraph or not.
        **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
            Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
            1.0 means token should be masked.

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **loss**: (`optional`, returned if both ``start_positions`` and ``end_positions`` are provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification losses.
thomwolf's avatar
thomwolf committed
708
        **start_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
709
710
            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
thomwolf's avatar
thomwolf committed
711
        **start_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
712
713
            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``
            Indices for the top config.start_n_top start token possibilities (beam-search).
thomwolf's avatar
thomwolf committed
714
        **end_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
715
716
            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
thomwolf's avatar
thomwolf committed
717
        **end_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
718
719
            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
thomwolf's avatar
thomwolf committed
720
        **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
721
722
            ``torch.FloatTensor`` of shape ``(batch_size,)``
            Log probabilities for the ``is_impossible`` label of the answers.
thomwolf's avatar
thomwolf committed
723
724
725
726
727
728
729
730
731
732
733
734
735
736
    """
    def __init__(self, config):
        super(SQuADHead, self).__init__()
        self.start_n_top = config.start_n_top
        self.end_n_top = config.end_n_top

        self.start_logits = PoolerStartLogits(config)
        self.end_logits = PoolerEndLogits(config)
        self.answer_class = PoolerAnswerClass(config)

    def forward(self, hidden_states, start_positions=None, end_positions=None,
                cls_index=None, is_impossible=None, p_mask=None):
        outputs = ()

thomwolf's avatar
thomwolf committed
737
        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
thomwolf's avatar
thomwolf committed
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760

        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, let's remove the dimension added by batch splitting
            for x in (start_positions, end_positions, cls_index, is_impossible):
                if x is not None and x.dim() > 1:
                    x.squeeze_(-1)

            # during training, compute the end logits based on the ground truth of the start position
            end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)

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

            if cls_index is not None and is_impossible is not None:
                # Predict answerability from the representation of CLS and START
                cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
                loss_fct_cls = nn.BCEWithLogitsLoss()
                cls_loss = loss_fct_cls(cls_logits, is_impossible)

                # note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
                total_loss += cls_loss * 0.5
761
762

            outputs = (total_loss,) + outputs
thomwolf's avatar
thomwolf committed
763
764
765
766
767
768
769

        else:
            # during inference, compute the end logits based on beam search
            bsz, slen, hsz = hidden_states.size()
            start_log_probs = F.softmax(start_logits, dim=-1) # shape (bsz, slen)

            start_top_log_probs, start_top_index = torch.topk(start_log_probs, self.start_n_top, dim=-1) # shape (bsz, start_n_top)
770
771
            start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz) # shape (bsz, start_n_top, hsz)
            start_states = torch.gather(hidden_states, -2, start_top_index_exp) # shape (bsz, start_n_top, hsz)
thomwolf's avatar
thomwolf committed
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
            start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1) # shape (bsz, slen, start_n_top, hsz)

            hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(start_states) # shape (bsz, slen, start_n_top, hsz)
            p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
            end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
            end_log_probs = F.softmax(end_logits, dim=1) # shape (bsz, slen, start_n_top)

            end_top_log_probs, end_top_index = torch.topk(end_log_probs, self.end_n_top, dim=1) # shape (bsz, end_n_top, start_n_top)
            end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
            end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)

            start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)
            cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)

            outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits) + outputs

        # return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
789
        # or (if labels are provided) (total_loss,)
thomwolf's avatar
thomwolf committed
790
791
792
793
        return outputs


class SequenceSummary(nn.Module):
thomwolf's avatar
thomwolf committed
794
    r""" Compute a single vector summary of a sequence hidden states according to various possibilities:
thomwolf's avatar
thomwolf committed
795
796
797
798
799
        Args of the config class:
            summary_type:
                - 'last' => [default] take the last token hidden state (like XLNet)
                - 'first' => take the first token hidden state (like Bert)
                - 'mean' => take the mean of all tokens hidden states
thomwolf's avatar
thomwolf committed
800
                - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2)
thomwolf's avatar
thomwolf committed
801
802
                - 'attn' => Not implemented now, use multi-head attention
            summary_use_proj: Add a projection after the vector extraction
803
            summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False.
804
            summary_activation: 'tanh' => add a tanh activation to the output, Other => no activation. Default
805
806
            summary_first_dropout: Add a dropout before the projection and activation
            summary_last_dropout: Add a dropout after the projection and activation
thomwolf's avatar
thomwolf committed
807
808
    """
    def __init__(self, config):
thomwolf's avatar
thomwolf committed
809
810
811
        super(SequenceSummary, self).__init__()

        self.summary_type = config.summary_type if hasattr(config, 'summary_use_proj') else 'last'
thomwolf's avatar
thomwolf committed
812
        if self.summary_type == 'attn':
thomwolf's avatar
thomwolf committed
813
814
815
816
817
            # We should use a standard multi-head attention module with absolute positional embedding for that.
            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
            raise NotImplementedError

thomwolf's avatar
thomwolf committed
818
        self.summary = Identity()
thomwolf's avatar
thomwolf committed
819
        if hasattr(config, 'summary_use_proj') and config.summary_use_proj:
820
821
            if hasattr(config, 'summary_proj_to_labels') and config.summary_proj_to_labels and config.num_labels > 0:
                num_classes = config.num_labels
thomwolf's avatar
thomwolf committed
822
823
824
825
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

thomwolf's avatar
thomwolf committed
826
        self.activation = Identity()
thomwolf's avatar
thomwolf committed
827
828
829
        if hasattr(config, 'summary_activation') and config.summary_activation == 'tanh':
            self.activation = nn.Tanh()

thomwolf's avatar
thomwolf committed
830
        self.first_dropout = Identity()
831
832
833
        if hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0:
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

thomwolf's avatar
thomwolf committed
834
        self.last_dropout = Identity()
835
836
        if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0:
            self.last_dropout = nn.Dropout(config.summary_last_dropout)
thomwolf's avatar
thomwolf committed
837

thomwolf's avatar
thomwolf committed
838
    def forward(self, hidden_states, cls_index=None):
thomwolf's avatar
thomwolf committed
839
        """ hidden_states: float Tensor in shape [bsz, seq_len, hidden_size], the hidden-states of the last layer.
thomwolf's avatar
thomwolf committed
840
            cls_index: [optional] position of the classification token if summary_type == 'cls_index',
thomwolf's avatar
thomwolf committed
841
                shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states.
thomwolf's avatar
thomwolf committed
842
                if summary_type == 'cls_index' and cls_index is None:
thomwolf's avatar
thomwolf committed
843
844
845
846
847
848
849
850
                    we take the last token of the sequence as classification token
        """
        if self.summary_type == 'last':
            output = hidden_states[:, -1]
        elif self.summary_type == 'first':
            output = hidden_states[:, 0]
        elif self.summary_type == 'mean':
            output = hidden_states.mean(dim=1)
thomwolf's avatar
thomwolf committed
851
852
853
        elif self.summary_type == 'cls_index':
            if cls_index is None:
                cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2]-1, dtype=torch.long)
thomwolf's avatar
thomwolf committed
854
            else:
thomwolf's avatar
thomwolf committed
855
856
857
858
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
                cls_index = cls_index.expand((-1,) * (cls_index.dim()-1) + (hidden_states.size(-1),))
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
            output = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, XX, hidden_size)
thomwolf's avatar
thomwolf committed
859
860
861
        elif self.summary_type == 'attn':
            raise NotImplementedError

862
        output = self.first_dropout(output)
thomwolf's avatar
thomwolf committed
863
864
        output = self.summary(output)
        output = self.activation(output)
865
        output = self.last_dropout(output)
thomwolf's avatar
thomwolf committed
866
867
868
869

        return output


870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
def prune_linear_layer(layer, index, dim=0):
    """ Prune a linear layer (a model parameters) to keep only entries in index.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    index = index.to(layer.weight.device)
    W = layer.weight.index_select(dim, index).clone().detach()
    if layer.bias is not None:
        if dim == 1:
            b = layer.bias.clone().detach()
        else:
            b = layer.bias[index].clone().detach()
    new_size = list(layer.weight.size())
    new_size[dim] = len(index)
    new_layer = nn.Linear(new_size[1], new_size[0], bias=layer.bias is not None).to(layer.weight.device)
    new_layer.weight.requires_grad = False
    new_layer.weight.copy_(W.contiguous())
    new_layer.weight.requires_grad = True
    if layer.bias is not None:
        new_layer.bias.requires_grad = False
        new_layer.bias.copy_(b.contiguous())
        new_layer.bias.requires_grad = True
    return new_layer


def prune_conv1d_layer(layer, index, dim=1):
    """ Prune a Conv1D layer (a model parameters) to keep only entries in index.
        A Conv1D work as a Linear layer (see e.g. BERT) but the weights are transposed.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    index = index.to(layer.weight.device)
    W = layer.weight.index_select(dim, index).clone().detach()
    if dim == 0:
        b = layer.bias.clone().detach()
    else:
        b = layer.bias[index].clone().detach()
    new_size = list(layer.weight.size())
    new_size[dim] = len(index)
    new_layer = Conv1D(new_size[1], new_size[0]).to(layer.weight.device)
    new_layer.weight.requires_grad = False
    new_layer.weight.copy_(W.contiguous())
    new_layer.weight.requires_grad = True
    new_layer.bias.requires_grad = False
    new_layer.bias.copy_(b.contiguous())
    new_layer.bias.requires_grad = True
    return new_layer
917
918
919
920
921
922
923
924
925
926
927
928
929


def prune_layer(layer, index, dim=None):
    """ Prune a Conv1D or nn.Linear layer (a model parameters) to keep only entries in index.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    if isinstance(layer, nn.Linear):
        return prune_linear_layer(layer, index, dim=0 if dim is None else dim)
    elif isinstance(layer, Conv1D):
        return prune_conv1d_layer(layer, index, dim=1 if dim is None else dim)
    else:
        raise ValueError("Can't prune layer of class {}".format(layer.__class__))