modeling_utils.py 85.1 KB
Newer Older
1
# coding=utf-8
thomwolf's avatar
thomwolf committed
2
# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# 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.

Patrick von Platen's avatar
Patrick von Platen committed
17
import inspect
18
import os
19
import re
20
import warnings
21
from dataclasses import dataclass
22
from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union
23
24

import torch
25
from torch import Tensor, device, dtype, nn
26
27
from torch.nn import CrossEntropyLoss
from torch.nn import functional as F
28

29
from .activations import get_activation
30
from .configuration_utils import PretrainedConfig
31
from .file_utils import (
Aymeric Augustin's avatar
Aymeric Augustin committed
32
    DUMMY_INPUTS,
33
34
35
    TF2_WEIGHTS_NAME,
    TF_WEIGHTS_NAME,
    WEIGHTS_NAME,
36
    ModelOutput,
37
38
39
    cached_path,
    hf_bucket_url,
    is_remote_url,
40
    is_torch_tpu_available,
Sylvain Gugger's avatar
Sylvain Gugger committed
41
    replace_return_docstrings,
42
)
43
from .generation_utils import GenerationMixin
Lysandre Debut's avatar
Lysandre Debut committed
44
from .utils import logging
45

Aymeric Augustin's avatar
Aymeric Augustin committed
46

Lysandre Debut's avatar
Lysandre Debut committed
47
logger = logging.get_logger(__name__)
48

thomwolf's avatar
thomwolf committed
49
50
51
52
53
try:
    from torch.nn import Identity
except ImportError:
    # Older PyTorch compatibility
    class Identity(nn.Module):
Lysandre's avatar
Lysandre committed
54
        r"""A placeholder identity operator that is argument-insensitive."""
55

thomwolf's avatar
thomwolf committed
56
        def __init__(self, *args, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
57
            super().__init__()
thomwolf's avatar
thomwolf committed
58
59
60
61

        def forward(self, input):
            return input

62

63
def find_pruneable_heads_and_indices(
Sylvain Gugger's avatar
Sylvain Gugger committed
64
65
66
67
68
69
70
71
72
73
74
75
76
77
    heads: List[int], n_heads: int, head_size: int, already_pruned_heads: Set[int]
) -> Tuple[Set[int], torch.LongTensor]:
    """
    Finds the heads and their indices taking :obj:`already_pruned_heads` into account.

    Args:
        heads (:obj:`List[int]`): List of the indices of heads to prune.
        n_heads (:obj:`int`): The number of heads in the model.
        head_size (:obj:`int`): The size of each head.
        already_pruned_heads (:obj:`Set[int]`): A set of already pruned heads.

    Returns:
        :obj:`Tuple[Set[int], torch.LongTensor]`: A tuple with the remaining heads and their corresponding indices.
    """
78
79
80
81
82
83
84
85
86
87
88
    mask = torch.ones(n_heads, head_size)
    heads = set(heads) - already_pruned_heads  # Convert to set and remove already pruned heads
    for head in heads:
        # Compute how many pruned heads are before the head and move the index accordingly
        head = head - sum(1 if h < head else 0 for h in already_pruned_heads)
        mask[head] = 0
    mask = mask.view(-1).contiguous().eq(1)
    index: torch.LongTensor = torch.arange(len(mask))[mask].long()
    return heads, index


Lysandre Debut's avatar
Lysandre Debut committed
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
def get_parameter_device(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    try:
        return next(parameter.parameters()).device
    except StopIteration:
        # For nn.DataParallel compatibility in PyTorch 1.5

        def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
            return tuples

        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
        first_tuple = next(gen)
        return first_tuple[1].device


def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
    try:
        return next(parameter.parameters()).dtype
    except StopIteration:
        # For nn.DataParallel compatibility in PyTorch 1.5

        def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
            return tuples

        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
        first_tuple = next(gen)
        return first_tuple[1].dtype


119
class ModuleUtilsMixin:
Julien Chaumond's avatar
Julien Chaumond committed
120
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
121
    A few utilities for :obj:`torch.nn.Modules`, to be used as a mixin.
Julien Chaumond's avatar
Julien Chaumond committed
122
123
    """

124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
    @staticmethod
    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
        try:
            import psutil
        except (ImportError):
            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")

        process = psutil.Process(os.getpid())
        mem = process.memory_info()
        module.mem_rss_pre_forward = mem.rss
        return None

    @staticmethod
    def _hook_rss_memory_post_forward(module, *args, **kwargs):
        try:
            import psutil
        except (ImportError):
            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")

        process = psutil.Process(os.getpid())
        mem = process.memory_info()
        module.mem_rss_post_forward = mem.rss
        mem_rss_diff = module.mem_rss_post_forward - module.mem_rss_pre_forward
        module.mem_rss_diff = mem_rss_diff + (module.mem_rss_diff if hasattr(module, "mem_rss_diff") else 0)
        return None

    def add_memory_hooks(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
151
152
153
154
155
        """
        Add a memory hook before and after each sub-module forward pass to record increase in memory consumption.

        Increase in memory consumption is stored in a :obj:`mem_rss_diff` attribute for each module and can be reset to
        zero with :obj:`model.reset_memory_hooks_state()`.
156
157
158
159
160
161
162
        """
        for module in self.modules():
            module.register_forward_pre_hook(self._hook_rss_memory_pre_forward)
            module.register_forward_hook(self._hook_rss_memory_post_forward)
        self.reset_memory_hooks_state()

    def reset_memory_hooks_state(self):
Sylvain Gugger's avatar
Sylvain Gugger committed
163
164
165
166
        """
        Reset the :obj:`mem_rss_diff` attribute of each module (see
        :func:`~transformers.modeling_utils.ModuleUtilsMixin.add_memory_hooks`).
        """
167
168
169
170
171
        for module in self.modules():
            module.mem_rss_diff = 0
            module.mem_rss_post_forward = 0
            module.mem_rss_pre_forward = 0

172
    @property
173
    def device(self) -> device:
174
        """
175
176
        :obj:`torch.device`: The device on which the module is (assuming that all the module parameters are on the same
        device).
177
        """
Lysandre Debut's avatar
Lysandre Debut committed
178
        return get_parameter_device(self)
179

180
181
    @property
    def dtype(self) -> dtype:
182
        """
183
        :obj:`torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
184
        """
Lysandre Debut's avatar
Lysandre Debut committed
185
        return get_parameter_dtype(self)
186
187

    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
Sylvain Gugger's avatar
Sylvain Gugger committed
188
189
190
191
192
193
194
195
196
        """
        Invert an attention mask (e.g., switches 0. and 1.).

        Args:
            encoder_attention_mask (:obj:`torch.Tensor`): An attention mask.

        Returns:
            :obj:`torch.Tensor`: The inverted attention mask.
        """
197
198
199
200
201
202
203
204
205
206
        if encoder_attention_mask.dim() == 3:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
        if encoder_attention_mask.dim() == 2:
            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow
        # /transformer/transformer_layers.py#L270
        # encoder_extended_attention_mask = (encoder_extended_attention_mask ==
        # encoder_extended_attention_mask.transpose(-1, -2))
        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
207
208
209
210
211
212
213
214
215
216
217
218

        if self.dtype == torch.float16:
            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e4
        elif self.dtype == torch.float32:
            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
        else:
            raise ValueError(
                "{} not recognized. `dtype` should be set to either `torch.float32` or `torch.float16`".format(
                    self.dtype
                )
            )

219
220
        return encoder_extended_attention_mask

Sylvain Gugger's avatar
Sylvain Gugger committed
221
222
223
    def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple[int], device: device) -> Tensor:
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
224
225

        Arguments:
Sylvain Gugger's avatar
Sylvain Gugger committed
226
227
228
229
230
231
            attention_mask (:obj:`torch.Tensor`):
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
            input_shape (:obj:`Tuple[int]`):
                The shape of the input to the model.
            device: (:obj:`torch.device`):
                The device of the input to the model.
232
233

        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
234
            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
235
236
237
238
239
240
241
242
243
244
245
246
247
        """
        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        if attention_mask.dim() == 3:
            extended_attention_mask = attention_mask[:, None, :, :]
        elif attention_mask.dim() == 2:
            # Provided a padding mask of dimensions [batch_size, seq_length]
            # - if the model is a decoder, apply a causal mask in addition to the padding mask
            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
            if self.config.is_decoder:
                batch_size, seq_length = input_shape
                seq_ids = torch.arange(seq_length, device=device)
                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
248
                # in case past_key_values are used we need to add a prefix ones mask to the causal mask
Patrick von Platen's avatar
Patrick von Platen committed
249
250
251
                # causal and attention masks must have same type with pytorch version < 1.3
                causal_mask = causal_mask.to(attention_mask.dtype)

252
253
254
                if causal_mask.shape[1] < attention_mask.shape[1]:
                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
                    causal_mask = torch.cat(
Patrick von Platen's avatar
Patrick von Platen committed
255
256
257
258
259
260
261
                        [
                            torch.ones(
                                (batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype
                            ),
                            causal_mask,
                        ],
                        axis=-1,
262
263
                    )

264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
            else:
                extended_attention_mask = attention_mask[:, None, None, :]
        else:
            raise ValueError(
                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
                    input_shape, attention_mask.shape
                )
            )

        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
        # masked positions, this operation will create a tensor which is 0.0 for
        # positions we want to attend and -10000.0 for masked positions.
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
        return extended_attention_mask

Sylvain Gugger's avatar
Sylvain Gugger committed
283
284
285
    def get_head_mask(
        self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
    ) -> Tensor:
286
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
287
288
289
290
291
292
293
294
295
296
        Prepare the head mask if needed.

        Args:
            head_mask (:obj:`torch.Tensor` with shape :obj:`[num_heads]` or :obj:`[num_hidden_layers x num_heads]`, `optional`):
                The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
            num_hidden_layers (:obj:`int`):
                The number of hidden layers in the model.
            is_attention_chunked: (:obj:`bool`, `optional, defaults to :obj:`False`):
                Whether or not the attentions scores are computed by chunks or not.

297
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
298
299
            :obj:`torch.Tensor` with shape :obj:`[num_hidden_layers x batch x num_heads x seq_length x seq_length]` or
            list with :obj:`[None]` for each layer.
300
301
302
        """
        if head_mask is not None:
            head_mask = self._convert_head_mask_to_5d(head_mask, num_hidden_layers)
Patrick von Platen's avatar
Patrick von Platen committed
303
304
            if is_attention_chunked is True:
                head_mask = head_mask.unsqueeze(-1)
305
306
307
308
309
310
311
312
313
314
315
316
317
        else:
            head_mask = [None] * num_hidden_layers

        return head_mask

    def _convert_head_mask_to_5d(self, head_mask, num_hidden_layers):
        """-> [num_hidden_layers x batch x num_heads x seq_length x seq_length]"""
        if head_mask.dim() == 1:
            head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
            head_mask = head_mask.expand(num_hidden_layers, -1, -1, -1, -1)
        elif head_mask.dim() == 2:
            head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
        assert head_mask.dim() == 5, f"head_mask.dim != 5, instead {head_mask.dim()}"
318
        head_mask = head_mask.to(dtype=self.dtype)  # switch to float if need + fp16 compatibility
319
320
        return head_mask

321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
    def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
        """
        Get number of (optionally, trainable or non-embeddings) parameters in the module.

        Args:
            only_trainable (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to return only the number of trainable parameters

            exclude_embeddings (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to return only the number of non-embeddings parameters

        Returns:
            :obj:`int`: The number of parameters.
        """

        def parameter_filter(x):
            return (x.requires_grad or not only_trainable) and not (
                isinstance(x, torch.nn.Embedding) and exclude_embeddings
            )

        params = filter(parameter_filter, self.parameters()) if only_trainable else self.parameters()
        return sum(p.numel() for p in params)

    def estimate_tokens(self, input_dict: Dict[str, Union[torch.Tensor, Any]]) -> int:
        """
        Helper function to estimate the total number of tokens from the model inputs.

        Args:
            inputs (:obj:`dict`): The model inputs.

        Returns:
            :obj:`int`: The total number of tokens.
        """
        token_inputs = [tensor for key, tensor in input_dict.items() if "input" in key]
        if token_inputs:
            return sum([token_input.numel() for token_input in token_inputs])
        else:
            warnings.warn(
                "Could not estimate the number of tokens of the input, floating-point operations will not be computed"
            )
            return 0

    def floating_point_ops(
        self, input_dict: Dict[str, Union[torch.Tensor, Any]], exclude_embeddings: bool = True
    ) -> int:
        """
        Get number of (optionally, non-embeddings) floating-point operations for the forward and backward passes of a
        batch with this transformer model. Default approximation neglects the quadratic dependency on the number of
Sylvain Gugger's avatar
Sylvain Gugger committed
369
        tokens (valid if :obj:`12 * d_model << sequence_length`) as laid out in `this paper
370
        <https://arxiv.org/pdf/2001.08361.pdf>`__ section 2.1. Should be overridden for transformers with parameter
Sylvain Gugger's avatar
Sylvain Gugger committed
371
        re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388

        Args:
            batch_size (:obj:`int`):
                The batch size for the forward pass.

            sequence_length (:obj:`int`):
                The number of tokens in each line of the batch.

            exclude_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
                Whether or not to count embedding and softmax operations.

        Returns:
            :obj:`int`: The number of floating-point operations.
        """

        return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)

Julien Chaumond's avatar
Julien Chaumond committed
389

390
class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
391
392
    r"""
    Base class for all models.
393

394
395
    :class:`~transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods
    for loading, downloading and saving models as well as a few methods common to all models to:
396

397
398
        * resize the input embeddings,
        * prune heads in the self-attention heads.
399

400
    Class attributes (overridden by derived classes):
Sylvain Gugger's avatar
Sylvain Gugger committed
401

402
403
        - **config_class** (:class:`~transformers.PretrainedConfig`) -- A subclass of
          :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
Sylvain Gugger's avatar
Sylvain Gugger committed
404
405
        - **load_tf_weights** (:obj:`Callable`) -- A python `method` for loading a TensorFlow checkpoint in a PyTorch
          model, taking as arguments:
406

407
408
            - **model** (:class:`~transformers.PreTrainedModel`) -- An instance of the model on which to load the
              TensorFlow checkpoint.
Sylvain Gugger's avatar
Sylvain Gugger committed
409
410
            - **config** (:class:`~transformers.PreTrainedConfig`) -- An instance of the configuration associated to
              the model.
411
412
413
414
            - **path** (:obj:`str`) -- A path to the TensorFlow checkpoint.

        - **base_model_prefix** (:obj:`str`) -- 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.
415
        - **is_parallelizable** (:obj:`bool`) -- A flag indicating whether this model supports model parallelization.
416
    """
417
    config_class = None
418
    base_model_prefix = ""
419
420
421
422
423
424
425
426
427
    # a list of re pattern of tensor names to ignore from the model when loading the model weights
    # (and avoid unnecessary warnings).
    _keys_to_ignore_on_load_missing = None
    # a list of re pattern of tensor names to ignore from the weights when loading the model weights
    # (and avoid unnecessary warnings).
    _keys_to_ignore_on_load_unexpected = None
    # a list of of tensor names to ignore when saving the model (useful for keys that aren't
    # trained, but which are deterministic)
    _keys_to_ignore_on_save = None
428

429
430
    is_parallelizable = False

431
    @property
432
    def dummy_inputs(self) -> Dict[str, torch.Tensor]:
433
434
        """
        :obj:`Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
435
        """
436
        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
437

438
    def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
Julien Chaumond's avatar
Julien Chaumond committed
439
        super().__init__()
440
441
442
443
444
445
        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__
446
447
                )
            )
448
        # Save config and origin of the pretrained weights if given in model
449
        self.config = config
450
        self.name_or_path = config.name_or_path
451

452
    @property
453
454
455
456
    def base_model(self) -> nn.Module:
        """
        :obj:`torch.nn.Module`: The main body of the model.
        """
457
        return getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
458

459
    def get_input_embeddings(self) -> nn.Module:
460
461
462
463
        """
        Returns the model's input embeddings.

        Returns:
464
            :obj:`nn.Module`: A torch module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
465
        """
466
        base_model = getattr(self, self.base_model_prefix, self)
thomwolf's avatar
thomwolf committed
467
468
469
470
        if base_model is not self:
            return base_model.get_input_embeddings()
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
471

472
    def set_input_embeddings(self, value: nn.Module):
473
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
474
        Set model's input embeddings.
475
476

        Args:
477
            value (:obj:`nn.Module`): A module mapping vocabulary to hidden states.
thomwolf's avatar
thomwolf committed
478
479
480
481
482
483
        """
        base_model = getattr(self, self.base_model_prefix, self)
        if base_model is not self:
            base_model.set_input_embeddings(value)
        else:
            raise NotImplementedError
thomwolf's avatar
thomwolf committed
484

485
    def get_output_embeddings(self) -> nn.Module:
486
487
488
489
        """
        Returns the model's output embeddings.

        Returns:
490
            :obj:`nn.Module`: A torch module mapping hidden states to vocabulary.
thomwolf's avatar
thomwolf committed
491
        """
492
        return None  # Overwrite for models with output embeddings
thomwolf's avatar
thomwolf committed
493

494
    def tie_weights(self):
495
496
        """
        Tie the weights between the input embeddings and the output embeddings.
497
498

        If the :obj:`torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning
499
        the weights instead.
thomwolf's avatar
thomwolf committed
500
        """
thomwolf's avatar
thomwolf committed
501
        output_embeddings = self.get_output_embeddings()
502
        if output_embeddings is not None and self.config.tie_word_embeddings:
thomwolf's avatar
thomwolf committed
503
            self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())
thomwolf's avatar
thomwolf committed
504

505
        if self.config.is_encoder_decoder and self.config.tie_encoder_decoder:
Weizhen's avatar
Weizhen committed
506
507
            if hasattr(self, self.base_model_prefix):
                self = getattr(self, self.base_model_prefix)
508
509
510
511
512
            self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)

    @staticmethod
    def _tie_encoder_decoder_weights(encoder: nn.Module, decoder: nn.Module, base_model_prefix: str):
        uninitialized_encoder_weights: List[str] = []
Weizhen's avatar
Weizhen committed
513
514
515
516
        if decoder.__class__ != encoder.__class__:
            logger.info(
                f"{decoder.__class__} and {encoder.__class__} are not equal. In this case make sure that all encoder weights are correctly initialized."
            )
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548

        def tie_encoder_to_decoder_recursively(
            decoder_pointer: nn.Module,
            encoder_pointer: nn.Module,
            module_name: str,
            uninitialized_encoder_weights: List[str],
            depth=0,
        ):
            assert isinstance(decoder_pointer, nn.Module) and isinstance(
                encoder_pointer, nn.Module
            ), f"{decoder_pointer} and {encoder_pointer} have to be of type torch.nn.Module"
            if hasattr(decoder_pointer, "weight"):
                assert hasattr(encoder_pointer, "weight")
                encoder_pointer.weight = decoder_pointer.weight
                if hasattr(decoder_pointer, "bias"):
                    assert hasattr(encoder_pointer, "bias")
                    encoder_pointer.bias = decoder_pointer.bias
                return

            encoder_modules = encoder_pointer._modules
            decoder_modules = decoder_pointer._modules
            if len(decoder_modules) > 0:
                assert (
                    len(encoder_modules) > 0
                ), f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"

                all_encoder_weights = set([module_name + "/" + sub_name for sub_name in encoder_modules.keys()])
                encoder_layer_pos = 0
                for name, module in decoder_modules.items():
                    if name.isdigit():
                        encoder_name = str(int(name) + encoder_layer_pos)
                        decoder_name = name
Weizhen's avatar
Weizhen committed
549
550
551
                        if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
                            encoder_modules
                        ) != len(decoder_modules):
552
553
                            # this can happen if the name corresponds to the position in a list module list of layers
                            # in this case the decoder has added a cross-attention that the encoder does not have
554
                            # thus skip this step and subtract one layer pos from encoder
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
                            encoder_layer_pos -= 1
                            continue
                    elif name not in encoder_modules:
                        continue
                    elif depth > 500:
                        raise ValueError(
                            "Max depth of recursive function `tie_encoder_to_decoder` reached. It seems that there is a circular dependency between two or more `nn.Modules` of your model."
                        )
                    else:
                        decoder_name = encoder_name = name
                    tie_encoder_to_decoder_recursively(
                        decoder_modules[decoder_name],
                        encoder_modules[encoder_name],
                        module_name + "/" + name,
                        uninitialized_encoder_weights,
                        depth=depth + 1,
                    )
                    all_encoder_weights.remove(module_name + "/" + encoder_name)

                uninitialized_encoder_weights += list(all_encoder_weights)

        # tie weights recursively
        tie_encoder_to_decoder_recursively(decoder, encoder, base_model_prefix, uninitialized_encoder_weights)
        if len(uninitialized_encoder_weights) > 0:
            logger.warning(
                f"The following encoder weights were not tied to the decoder {uninitialized_encoder_weights}"
            )

583
    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
Lysandre's avatar
Lysandre committed
584
        """Tie or clone module weights depending of whether we are using TorchScript or not"""
thomwolf's avatar
thomwolf committed
585
        if self.config.torchscript:
586
            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
thomwolf's avatar
thomwolf committed
587
        else:
588
            output_embeddings.weight = input_embeddings.weight
thomwolf's avatar
thomwolf committed
589

Sam Shleifer's avatar
Sam Shleifer committed
590
        if getattr(output_embeddings, "bias", None) is not None:
591
592
            output_embeddings.bias.data = torch.nn.functional.pad(
                output_embeddings.bias.data,
Lysandre's avatar
Lysandre committed
593
594
595
596
                (
                    0,
                    output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
                ),
597
598
                "constant",
                0,
599
            )
600
        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
601
            output_embeddings.out_features = input_embeddings.num_embeddings
602

603
604
605
    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> torch.nn.Embedding:
        """
        Resizes input token embeddings matrix of the model if :obj:`new_num_tokens != config.vocab_size`.
606

607
        Takes care of tying weights embeddings afterwards if the model class has a :obj:`tie_weights()` method.
thomwolf's avatar
thomwolf committed
608

609
610
611
612
        Arguments:
            new_num_tokens (:obj:`int`, `optional`):
                The number of new 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 :obj:`None`,
613
                just returns a pointer to the input tokens :obj:`torch.nn.Embedding` module of the model without doing
614
615
616
617
                anything.

        Return:
            :obj:`torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
thomwolf's avatar
thomwolf committed
618
        """
619
        model_embeds = self._resize_token_embeddings(new_num_tokens)
thomwolf's avatar
thomwolf committed
620
621
        if new_num_tokens is None:
            return model_embeds
thomwolf's avatar
thomwolf committed
622
623
624

        # Update base model and current model config
        self.config.vocab_size = new_num_tokens
625
        self.vocab_size = new_num_tokens
thomwolf's avatar
thomwolf committed
626
627

        # Tie weights again if needed
628
        self.tie_weights()
thomwolf's avatar
thomwolf committed
629

thomwolf's avatar
thomwolf committed
630
631
        return model_embeds

632
    def _resize_token_embeddings(self, new_num_tokens):
thomwolf's avatar
thomwolf committed
633
634
635
        old_embeddings = self.get_input_embeddings()
        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
        self.set_input_embeddings(new_embeddings)
636
637
638
639
640
641
642

        # if word embeddings are not tied, make sure that lm head is resized as well
        if self.get_output_embeddings() is not None and not self.config.tie_word_embeddings:
            old_lm_head = self.get_output_embeddings()
            new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
            self.set_output_embeddings(new_lm_head)

thomwolf's avatar
thomwolf committed
643
        return self.get_input_embeddings()
644

645
646
647
    def _get_resized_embeddings(
        self, old_embeddings: torch.nn.Embedding, new_num_tokens: Optional[int] = None
    ) -> torch.nn.Embedding:
648
649
650
        """
        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
651
652

        Args:
653
            old_embeddings (:obj:`torch.nn.Embedding`):
654
                Old embeddings to be resized.
655
            new_num_tokens (:obj:`int`, `optional`):
656
                New number of tokens in the embedding matrix.
657
658
659

                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
                vectors from the end. If not provided or :obj:`None`, just returns a pointer to the input tokens
660
                :obj:`torch.nn.Embedding`` module of the model without doing anything.
661
662
663
664

        Return:
            :obj:`torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
            :obj:`new_num_tokens` is :obj:`None`
665
666
667
668
669
670
671
672
        """
        if new_num_tokens is None:
            return old_embeddings

        old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
        if old_num_tokens == new_num_tokens:
            return old_embeddings

673
674
675
676
677
678
        if not isinstance(old_embeddings, nn.Embedding):
            raise TypeError(
                f"Old embeddings are of type {type(old_embeddings)}, which is not an instance of {nn.Embedding}."
                f"You should either use a different resize function or make sure that `old_embeddings` are an instance of {nn.Embedding}."
            )

679
        # Build new embeddings
680
        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim).to(self.device)
681
682
683
684

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

685
        # Copy token embeddings from the previous weights
686
687
688
689
690
        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

691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
    def _get_resized_lm_head(
        self, old_lm_head: torch.nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
    ) -> torch.nn.Linear:
        """
        Build a resized Linear Module from a provided old Linear Module. Increasing the size will add newly initialized
        vectors at the end. Reducing the size will remove vectors from the end

        Args:
            old_lm_head (:obj:`torch.nn.Linear`):
                Old lm head liner layer to be resized.
            new_num_tokens (:obj:`int`, `optional`):
                New number of tokens in the linear 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 :obj:`None`, just returns a pointer to the input tokens
                :obj:`torch.nn.Linear`` module of the model without doing anything.
            transposed (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether ``old_lm_head`` is transposed or not. If True ``old_lm_head.size()`` is ``lm_head_dim,
                vocab_size`` else ``vocab_size, lm_head_dim``.

        Return:
            :obj:`torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if
            :obj:`new_num_tokens` is :obj:`None`
        """
        if new_num_tokens is None:
            return old_lm_head

        old_num_tokens, old_lm_head_dim = (
            old_lm_head.weight.size() if not transposed else old_lm_head.weight.t().size()
        )

        if old_num_tokens == new_num_tokens:
            return old_lm_head

        if not isinstance(old_lm_head, nn.Linear):
            raise TypeError(
                f"Old language model head is of type {type(old_lm_head)}, which is not an instance of {nn.Linear}."
                f"You should either use a different resize function or make sure that `old_embeddings` are an instance of {nn.Linear}."
            )

        # Build new lm head
        new_lm_head_shape = (old_lm_head_dim, new_num_tokens) if not transposed else (new_num_tokens, old_lm_head_dim)
        has_new_lm_head_bias = old_lm_head.bias is not None
        new_lm_head = nn.Linear(*new_lm_head_shape, bias=has_new_lm_head_bias).to(self.device)

        # initialize new lm head (in particular added tokens)
        self._init_weights(new_lm_head)

        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)

        # Copy old lm head weights to new lm head
        if not transposed:
            new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[:num_tokens_to_copy, :]
        else:
            new_lm_head.weight.data[:, :num_tokens_to_copy] = old_lm_head.weight.data[:, :num_tokens_to_copy]

        # Copy bias weights to new lm head
        if has_new_lm_head_bias:
            new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]

        return new_lm_head

753
    def init_weights(self):
754
755
756
        """
        Initializes and prunes weights if needed.
        """
757
758
759
760
761
762
763
        # Initialize weights
        self.apply(self._init_weights)

        # Prune heads if needed
        if self.config.pruned_heads:
            self.prune_heads(self.config.pruned_heads)

764
765
766
        # Tie weights if needed
        self.tie_weights()

767
768
769
    def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
        """
        Prunes heads of the base model.
770

771
772
        Arguments:
            heads_to_prune (:obj:`Dict[int, List[int]]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
773
774
775
                Dictionary with keys being selected layer indices (:obj:`int`) and associated values being the list of
                heads to prune in said layer (list of :obj:`int`). For instance {1: [0, 2], 2: [2, 3]} will prune heads
                0 and 2 on layer 1 and heads 2 and 3 on layer 2.
thomwolf's avatar
thomwolf committed
776
        """
777
        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
778
        for layer, heads in heads_to_prune.items():
779
780
781
            union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
            self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON

782
        self.base_model._prune_heads(heads_to_prune)
thomwolf's avatar
thomwolf committed
783

784
    def save_pretrained(self, save_directory: Union[str, os.PathLike]):
785
786
787
        """
        Save a model and its configuration file to a directory, so that it can be re-loaded using the
        `:func:`~transformers.PreTrainedModel.from_pretrained`` class method.
788

789
        Arguments:
790
            save_directory (:obj:`str` or :obj:`os.PathLike`):
791
                Directory to which to save. Will be created if it doesn't exist.
792
        """
793
794
795
796
        if os.path.isfile(save_directory):
            logger.error("Provided path ({}) should be a directory, not a file".format(save_directory))
            return
        os.makedirs(save_directory, exist_ok=True)
797

Julien Chaumond's avatar
Julien Chaumond committed
798
        # Only save the model itself if we are using distributed training
799
        model_to_save = self.module if hasattr(self, "module") else self
800

Julien Chaumond's avatar
Julien Chaumond committed
801
802
803
        # Attach architecture to the config
        model_to_save.config.architectures = [model_to_save.__class__.__name__]

804
805
806
        state_dict = model_to_save.state_dict()

        # Handle the case where some state_dict keys shouldn't be saved
807
808
        if self._keys_to_ignore_on_save is not None:
            state_dict = {k: v for k, v in state_dict.items() if k not in self._keys_to_ignore_on_save}
809

810
811
        # If we save using the predefined names, we can load using `from_pretrained`
        output_model_file = os.path.join(save_directory, WEIGHTS_NAME)
812

813
        if getattr(self.config, "xla_device", False) and is_torch_tpu_available():
814
815
816
817
818
819
            import torch_xla.core.xla_model as xm

            if xm.is_master_ordinal():
                # Save configuration file
                model_to_save.config.save_pretrained(save_directory)
            # xm.save takes care of saving only from master
820
            xm.save(state_dict, output_model_file)
821
822
        else:
            model_to_save.config.save_pretrained(save_directory)
823
            torch.save(state_dict, output_model_file)
824

thomwolf's avatar
thomwolf committed
825
        logger.info("Model weights saved in {}".format(output_model_file))
826

827
    @classmethod
828
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
829
830
        r"""
        Instantiate a pretrained pytorch model from a pre-trained model configuration.
831

Sylvain Gugger's avatar
Sylvain Gugger committed
832
833
        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()``.
834

835
836
837
        The warning `Weights from XXX not initialized from pretrained model` means that the weights of XXX do not come
        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
        task.
838

839
840
        The warning `Weights from XXX not used in YYY` means that the layer XXX is not used by YYY, therefore those
        weights are discarded.
841

842
        Parameters:
843
            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`, `optional`):
844
845
                Can be either:

846
847
848
                    - A string, the `model id` of a pretrained model hosted inside a model repo on huggingface.co.
                      Valid model ids can be located at the root-level, like ``bert-base-uncased``, or namespaced under
                      a user or organization name, like ``dbmdz/bert-base-german-cased``.
849
850
                    - A path to a `directory` containing model weights saved using
                      :func:`~transformers.PreTrainedModel.save_pretrained`, e.g., ``./my_model_directory/``.
Sylvain Gugger's avatar
Sylvain Gugger committed
851
                    - A path or url to a `tensorflow index checkpoint file` (e.g, ``./tf_model/model.ckpt.index``). In
852
853
854
855
856
857
858
                      this case, ``from_tf`` should be set to :obj:`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.
                    - :obj:`None` if you are both providing the configuration and state dictionary (resp. with keyword
                      arguments ``config`` and ``state_dict``).
            model_args (sequence of positional arguments, `optional`):
                All remaning positional arguments will be passed to the underlying model's ``__init__`` method.
859
            config (:obj:`Union[PretrainedConfig, str, os.PathLike]`, `optional`):
860
861
862
                Can be either:

                    - an instance of a class derived from :class:`~transformers.PretrainedConfig`,
863
                    - a string or path valid as input to :func:`~transformers.PretrainedConfig.from_pretrained`.
864
865
866
867

                Configuration for the model to use instead of an automatically loaded configuation. Configuration can
                be automatically loaded when:

868
869
                    - The model is a model provided by the library (loaded with the `model id` string of a pretrained
                      model).
870
                    - The model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded
871
872
                      by supplying the save directory.
                    - The model is loaded by supplying a local directory as ``pretrained_model_name_or_path`` and a
873
874
875
876
877
878
879
880
                      configuration JSON file named `config.json` is found in the directory.
            state_dict (:obj:`Dict[str, torch.Tensor]`, `optional`):
                A state dictionary to use instead of a state dictionary loaded from saved weights file.

                This option can be used if you want to create a model from a pretrained configuration but load your own
                weights. In this case though, you should check if using
                :func:`~transformers.PreTrainedModel.save_pretrained` and
                :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
881
            cache_dir (:obj:`Union[str, os.PathLike]`, `optional`):
882
883
884
885
886
887
888
889
890
891
892
                Path to a directory in which a downloaded pretrained model configuration should be cached if the
                standard cache should not be used.
            from_tf (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Load the model weights from a TensorFlow checkpoint save file (see docstring of
                ``pretrained_model_name_or_path`` argument).
            force_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
                cached versions if they exist.
            resume_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
                file exists.
Sylvain Gugger's avatar
Sylvain Gugger committed
893
            proxies (:obj:`Dict[str, str], `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
894
895
                A dictionary of proxy servers to use by protocol or endpoint, e.g., :obj:`{'http': 'foo.bar:3128',
                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
896
            output_loading_info(:obj:`bool`, `optional`, defaults to :obj:`False`):
Sylvain Gugger's avatar
Sylvain Gugger committed
897
                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
898
            local_files_only(:obj:`bool`, `optional`, defaults to :obj:`False`):
Stas Bekman's avatar
Stas Bekman committed
899
                Whether or not to only look at local files (i.e., do not try to download the model).
900
901
902
            use_auth_token (:obj:`str` or `bool`, `optional`):
                The token to use as HTTP bearer authorization for remote files. If :obj:`True`, will use the token
                generated when running :obj:`transformers-cli login` (stored in :obj:`~/.huggingface`).
Julien Chaumond's avatar
Julien Chaumond committed
903
904
905
906
            revision(:obj:`str`, `optional`, defaults to :obj:`"main"`):
                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
                git-based system for storing models and other artifacts on huggingface.co, so ``revision`` can be any
                identifier allowed by git.
907
            mirror(:obj:`str`, `optional`, defaults to :obj:`None`):
Sylvain Gugger's avatar
Sylvain Gugger committed
908
909
910
                Mirror source to accelerate downloads in China. If you are from China and have an accessibility
                problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
                Please refer to the mirror site for more information.
911
912
            kwargs (remaining dictionary of keyword arguments, `optional`):
                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
913
                :obj:`output_attentions=True`). Behaves differently depending on whether a ``config`` is provided or
914
915
916
917
918
919
920
921
922
923
                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:`~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.
924

925
926
927
928
        .. note::

            Passing :obj:`use_auth_token=True` is required when you want to use a private model.

929
        Examples::
thomwolf's avatar
thomwolf committed
930

931
            >>> from transformers import BertConfig, BertModel
932
            >>> # Download model and configuration from huggingface.co and cache.
933
934
935
936
937
938
939
940
941
            >>> model = BertModel.from_pretrained('bert-base-uncased')
            >>> # Model was saved using `save_pretrained('./test/saved_model/')` (for example purposes, not runnable).
            >>> model = BertModel.from_pretrained('./test/saved_model/')
            >>> # Update configuration during loading.
            >>> model = BertModel.from_pretrained('bert-base-uncased', output_attentions=True)
            >>> assert model.config.output_attentions == True
            >>> # Loading from a TF checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
            >>> 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)
942
        """
943
944
945
946
947
948
949
950
        config = kwargs.pop("config", None)
        state_dict = kwargs.pop("state_dict", None)
        cache_dir = kwargs.pop("cache_dir", None)
        from_tf = kwargs.pop("from_tf", False)
        force_download = kwargs.pop("force_download", False)
        resume_download = kwargs.pop("resume_download", False)
        proxies = kwargs.pop("proxies", None)
        output_loading_info = kwargs.pop("output_loading_info", False)
951
        local_files_only = kwargs.pop("local_files_only", False)
952
        use_auth_token = kwargs.pop("use_auth_token", None)
Julien Chaumond's avatar
Julien Chaumond committed
953
        revision = kwargs.pop("revision", None)
954
        mirror = kwargs.pop("mirror", None)
thomwolf's avatar
thomwolf committed
955

956
957
958
        # Load config if we don't provide a configuration
        if not isinstance(config, PretrainedConfig):
            config_path = config if config is not None else pretrained_model_name_or_path
959
            config, model_kwargs = cls.config_class.from_pretrained(
960
961
962
963
                config_path,
                *model_args,
                cache_dir=cache_dir,
                return_unused_kwargs=True,
964
                force_download=force_download,
965
                resume_download=resume_download,
966
                proxies=proxies,
967
                local_files_only=local_files_only,
968
                use_auth_token=use_auth_token,
Julien Chaumond's avatar
Julien Chaumond committed
969
                revision=revision,
970
                **kwargs,
971
972
973
            )
        else:
            model_kwargs = kwargs
974

thomwolf's avatar
thomwolf committed
975
        # Load model
thomwolf's avatar
thomwolf committed
976
        if pretrained_model_name_or_path is not None:
977
            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
978
            if os.path.isdir(pretrained_model_name_or_path):
thomwolf's avatar
thomwolf committed
979
                if from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")):
980
                    # Load from a TF 1.0 checkpoint in priority if from_tf
thomwolf's avatar
thomwolf committed
981
                    archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
thomwolf's avatar
thomwolf committed
982
                elif from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
983
                    # Load from a TF 2.0 checkpoint in priority if from_tf
thomwolf's avatar
thomwolf committed
984
985
986
                    archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
                    # Load from a PyTorch checkpoint
thomwolf's avatar
thomwolf committed
987
                    archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
thomwolf's avatar
thomwolf committed
988
                else:
989
990
                    raise EnvironmentError(
                        "Error no file named {} found in directory {} or `from_tf` set to False".format(
Patrick von Platen's avatar
Patrick von Platen committed
991
                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index"],
Patrick von Platen's avatar
Patrick von Platen committed
992
                            pretrained_model_name_or_path,
993
994
                        )
                    )
995
            elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
996
                archive_file = pretrained_model_name_or_path
997
            elif os.path.isfile(pretrained_model_name_or_path + ".index"):
998
999
1000
1001
1002
                assert (
                    from_tf
                ), "We found a TensorFlow checkpoint at {}, please set from_tf to True to load from this checkpoint".format(
                    pretrained_model_name_or_path + ".index"
                )
1003
                archive_file = pretrained_model_name_or_path + ".index"
1004
            else:
thomwolf's avatar
thomwolf committed
1005
                archive_file = hf_bucket_url(
Julien Chaumond's avatar
Julien Chaumond committed
1006
1007
                    pretrained_model_name_or_path,
                    filename=(TF2_WEIGHTS_NAME if from_tf else WEIGHTS_NAME),
Julien Chaumond's avatar
Julien Chaumond committed
1008
                    revision=revision,
1009
                    mirror=mirror,
thomwolf's avatar
thomwolf committed
1010
                )
1011

thomwolf's avatar
thomwolf committed
1012
            try:
1013
                # Load from URL or cache if already cached
1014
1015
1016
1017
1018
1019
                resolved_archive_file = cached_path(
                    archive_file,
                    cache_dir=cache_dir,
                    force_download=force_download,
                    proxies=proxies,
                    resume_download=resume_download,
1020
                    local_files_only=local_files_only,
1021
                    use_auth_token=use_auth_token,
1022
                )
Julien Chaumond's avatar
Julien Chaumond committed
1023
1024
            except EnvironmentError as err:
                logger.error(err)
1025
1026
1027
1028
1029
                msg = (
                    f"Can't load weights for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
                    f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
                    f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a file named one of {WEIGHTS_NAME}, {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME}.\n\n"
                )
thomwolf's avatar
thomwolf committed
1030
1031
                raise EnvironmentError(msg)

thomwolf's avatar
thomwolf committed
1032
1033
            if resolved_archive_file == archive_file:
                logger.info("loading weights file {}".format(archive_file))
1034
            else:
1035
                logger.info("loading weights file {} from cache at {}".format(archive_file, resolved_archive_file))
1036
        else:
thomwolf's avatar
thomwolf committed
1037
            resolved_archive_file = None
1038

1039
1040
        config.name_or_path = pretrained_model_name_or_path

1041
        # Instantiate model.
1042
        model = cls(config, *model_args, **model_kwargs)
thomwolf's avatar
thomwolf committed
1043

1044
        if state_dict is None and not from_tf:
1045
            try:
1046
                state_dict = torch.load(resolved_archive_file, map_location="cpu")
1047
            except Exception:
1048
                raise OSError(
1049
1050
                    f"Unable to load weights from pytorch checkpoint file for '{pretrained_model_name_or_path}' "
                    f"at '{resolved_archive_file}'"
1051
1052
                    "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. "
                )
1053

1054
1055
1056
        missing_keys = []
        unexpected_keys = []
        error_msgs = []
1057
1058

        if from_tf:
1059
            if resolved_archive_file.endswith(".index"):
1060
1061
1062
1063
1064
                # Load from a TensorFlow 1.X checkpoint - provided by original authors
                model = cls.load_tf_weights(model, config, resolved_archive_file[:-6])  # Remove the '.index'
            else:
                # Load from our TensorFlow 2.0 checkpoints
                try:
1065
                    from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
1066

1067
                    model = load_tf2_checkpoint_in_pytorch_model(model, resolved_archive_file, allow_missing_keys=True)
1068
                except ImportError:
1069
1070
1071
1072
                    logger.error(
                        "Loading a TensorFlow model in PyTorch, requires both PyTorch and TensorFlow to be installed. Please see "
                        "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
                    )
1073
                    raise
1074
1075
1076
1077
1078
1079
        else:
            # 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
1080
1081
1082
1083
                if "gamma" in key:
                    new_key = key.replace("gamma", "weight")
                if "beta" in key:
                    new_key = key.replace("beta", "bias")
1084
1085
1086
1087
1088
1089
1090
                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)

            # copy state_dict so _load_from_state_dict can modify it
1091
            metadata = getattr(state_dict, "_metadata", None)
1092
1093
1094
1095
            state_dict = state_dict.copy()
            if metadata is not None:
                state_dict._metadata = metadata

1096
1097
            # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
            # so we need to apply the function recursively.
Julien Chaumond's avatar
Julien Chaumond committed
1098
            def load(module: nn.Module, prefix=""):
1099
1100
                local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
                module._load_from_state_dict(
Lysandre's avatar
Lysandre committed
1101
1102
1103
1104
1105
1106
1107
                    state_dict,
                    prefix,
                    local_metadata,
                    True,
                    missing_keys,
                    unexpected_keys,
                    error_msgs,
1108
                )
1109
1110
                for name, child in module._modules.items():
                    if child is not None:
1111
                        load(child, prefix + name + ".")
1112
1113

            # Make sure we are able to load base models as well as derived models (with heads)
1114
            start_prefix = ""
1115
            model_to_load = model
1116
1117
            has_prefix_module = any(s.startswith(cls.base_model_prefix) for s in state_dict.keys())
            if not hasattr(model, cls.base_model_prefix) and has_prefix_module:
1118
                start_prefix = cls.base_model_prefix + "."
1119
            if hasattr(model, cls.base_model_prefix) and not has_prefix_module:
1120
1121
1122
                model_to_load = getattr(model, cls.base_model_prefix)

            load(model_to_load, prefix=start_prefix)
1123
1124
1125
1126
1127
1128
1129
1130

            if model.__class__.__name__ != model_to_load.__class__.__name__:
                base_model_state_dict = model_to_load.state_dict().keys()
                head_model_state_dict_without_base_prefix = [
                    key.split(cls.base_model_prefix + ".")[-1] for key in model.state_dict().keys()
                ]
                missing_keys.extend(head_model_state_dict_without_base_prefix - base_model_state_dict)

1131
1132
            # Some models may have keys that are not in the state by design, removing them before needlessly warning
            # the user.
1133
1134
            if cls._keys_to_ignore_on_load_missing is not None:
                for pat in cls._keys_to_ignore_on_load_missing:
1135
1136
                    missing_keys = [k for k in missing_keys if re.search(pat, k) is None]

1137
1138
            if cls._keys_to_ignore_on_load_unexpected is not None:
                for pat in cls._keys_to_ignore_on_load_unexpected:
1139
1140
                    unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]

1141
1142
1143
1144
1145
            if len(unexpected_keys) > 0:
                logger.warning(
                    f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when "
                    f"initializing {model.__class__.__name__}: {unexpected_keys}\n"
                    f"- This IS expected if you are initializing {model.__class__.__name__} from the checkpoint of a model trained on another task "
1146
                    f"or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n"
1147
1148
1149
1150
1151
                    f"- This IS NOT expected if you are initializing {model.__class__.__name__} from the checkpoint of a model that you expect "
                    f"to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
                )
            else:
                logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
1152
            if len(missing_keys) > 0:
1153
1154
1155
1156
                logger.warning(
                    f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at {pretrained_model_name_or_path} "
                    f"and are newly initialized: {missing_keys}\n"
                    f"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference."
1157
                )
1158
            else:
1159
                logger.info(
1160
                    f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at {pretrained_model_name_or_path}.\n"
Prajjwal Bhargava's avatar
Prajjwal Bhargava committed
1161
                    f"If your task is similar to the task the model of the checkpoint was trained on, "
1162
                    f"you can already use {model.__class__.__name__} for predictions without further training."
1163
                )
1164
            if len(error_msgs) > 0:
1165
1166
1167
1168
1169
                raise RuntimeError(
                    "Error(s) in loading state_dict for {}:\n\t{}".format(
                        model.__class__.__name__, "\n\t".join(error_msgs)
                    )
                )
1170
1171
        # make sure token embedding weights are still tied if needed
        model.tie_weights()
1172

1173
        # Set model in evaluation mode to deactivate DropOut modules by default
1174
1175
        model.eval()

thomwolf's avatar
thomwolf committed
1176
        if output_loading_info:
1177
1178
1179
1180
1181
            loading_info = {
                "missing_keys": missing_keys,
                "unexpected_keys": unexpected_keys,
                "error_msgs": error_msgs,
            }
thomwolf's avatar
thomwolf committed
1182
1183
            return model, loading_info

1184
        if hasattr(config, "xla_device") and config.xla_device and is_torch_tpu_available():
1185
1186
1187
            import torch_xla.core.xla_model as xm

            model = xm.send_cpu_data_to_device(model, xm.xla_device())
1188
            model.to(xm.xla_device())
1189

1190
1191
        return model

thomwolf's avatar
thomwolf committed
1192

thomwolf's avatar
thomwolf committed
1193
class Conv1D(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
    """
    1D-convolutional layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2).

    Basically works like a linear layer but the weights are transposed.

    Args:
        nf (:obj:`int`): The number of output features.
        nx (:obj:`int`): The number of input features.
    """

thomwolf's avatar
thomwolf committed
1204
    def __init__(self, nf, nx):
Julien Chaumond's avatar
Julien Chaumond committed
1205
        super().__init__()
thomwolf's avatar
thomwolf committed
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
        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
1219
class PoolerStartLogits(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
1220
1221
    """
    Compute SQuAD start logits from sequence hidden states.
1222

Sylvain Gugger's avatar
Sylvain Gugger committed
1223
1224
1225
1226
1227
1228
    Args:
        config (:class:`~transformers.PretrainedConfig`):
            The config used by the model, will be used to grab the :obj:`hidden_size` of the model.
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
1229
        super().__init__()
thomwolf's avatar
thomwolf committed
1230
1231
        self.dense = nn.Linear(config.hidden_size, 1)

Sylvain Gugger's avatar
Sylvain Gugger committed
1232
1233
1234
1235
1236
1237
1238
1239
    def forward(
        self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
    ) -> torch.FloatTensor:
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
                The final hidden states of the model.
            p_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1240
1241
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
Sylvain Gugger's avatar
Sylvain Gugger committed
1242
1243
1244

        Returns:
            :obj:`torch.FloatTensor`: The start logits for SQuAD.
thomwolf's avatar
thomwolf committed
1245
        """
thomwolf's avatar
thomwolf committed
1246
1247
1248
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
Lysandre Debut's avatar
Lysandre Debut committed
1249
            if get_parameter_dtype(self) == torch.float16:
1250
1251
1252
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
1253
1254
1255
1256
1257
1258

        return x


class PoolerEndLogits(nn.Module):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
1259
    Compute SQuAD end logits from sequence hidden states.
1260

Sylvain Gugger's avatar
Sylvain Gugger committed
1261
1262
1263
1264
1265
1266
1267
    Args:
        config (:class:`~transformers.PretrainedConfig`):
            The config used by the model, will be used to grab the :obj:`hidden_size` of the model and the
            :obj:`layer_norm_eps` to use.
    """

    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
1268
        super().__init__()
thomwolf's avatar
thomwolf committed
1269
1270
1271
1272
1273
        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)

Sylvain Gugger's avatar
Sylvain Gugger committed
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
    def forward(
        self,
        hidden_states: torch.FloatTensor,
        start_states: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        p_mask: Optional[torch.FloatTensor] = None,
    ) -> torch.FloatTensor:
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
                The final hidden states of the model.
            start_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`, `optional`):
                The hidden states of the first tokens for the labeled span.
            start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                The position of the first token for the labeled span.
            p_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1290
1291
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
Sylvain Gugger's avatar
Sylvain Gugger committed
1292
1293
1294
1295
1296
1297
1298
1299

        .. note::

            One of ``start_states`` or ``start_positions`` should be not obj:`None`. If both are set,
            ``start_positions`` overrides ``start_states``.

        Returns:
            :obj:`torch.FloatTensor`: The end logits for SQuAD.
thomwolf's avatar
thomwolf committed
1300
        """
1301
1302
1303
        assert (
            start_states is not None or start_positions is not None
        ), "One of start_states, start_positions should be not None"
thomwolf's avatar
thomwolf committed
1304
        if start_positions is not None:
1305
            slen, hsz = hidden_states.shape[-2:]
1306
1307
1308
            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)
thomwolf's avatar
thomwolf committed
1309
1310
1311
1312
1313
1314
1315

        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:
Lysandre Debut's avatar
Lysandre Debut committed
1316
            if get_parameter_dtype(self) == torch.float16:
1317
1318
1319
                x = x * (1 - p_mask) - 65500 * p_mask
            else:
                x = x * (1 - p_mask) - 1e30 * p_mask
thomwolf's avatar
thomwolf committed
1320
1321
1322
1323
1324

        return x


class PoolerAnswerClass(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
1325
1326
1327
1328
1329
1330
1331
    """
    Compute SQuAD 2.0 answer class from classification and start tokens hidden states.

    Args:
        config (:class:`~transformers.PretrainedConfig`):
            The config used by the model, will be used to grab the :obj:`hidden_size` of the model.
    """
1332

thomwolf's avatar
thomwolf committed
1333
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
1334
        super().__init__()
thomwolf's avatar
thomwolf committed
1335
1336
1337
1338
        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)

Sylvain Gugger's avatar
Sylvain Gugger committed
1339
1340
1341
1342
1343
1344
1345
    def forward(
        self,
        hidden_states: torch.FloatTensor,
        start_states: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        cls_index: Optional[torch.LongTensor] = None,
    ) -> torch.FloatTensor:
1346
1347
        """
        Args:
Sylvain Gugger's avatar
Sylvain Gugger committed
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
                The final hidden states of the model.
            start_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`, `optional`):
                The hidden states of the first tokens for the labeled span.
            start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                The position of the first token for the labeled span.
            cls_index (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                Position of the CLS token for each sentence in the batch. If :obj:`None`, takes the last token.

        .. note::

            One of ``start_states`` or ``start_positions`` should be not obj:`None`. If both are set,
            ``start_positions`` overrides ``start_states``.

        Returns:
            :obj:`torch.FloatTensor`: The SQuAD 2.0 answer class.
thomwolf's avatar
thomwolf committed
1364
        """
Sylvain Gugger's avatar
Sylvain Gugger committed
1365
        # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
1366
        hsz = hidden_states.shape[-1]
1367
1368
1369
        assert (
            start_states is not None or start_positions is not None
        ), "One of start_states, start_positions should be not None"
thomwolf's avatar
thomwolf committed
1370
        if start_positions is not None:
1371
1372
            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)
thomwolf's avatar
thomwolf committed
1373
1374

        if cls_index is not None:
1375
1376
            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)
thomwolf's avatar
thomwolf committed
1377
        else:
1378
            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
thomwolf's avatar
thomwolf committed
1379
1380
1381
1382
1383
1384
1385
1386

        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


1387
1388
1389
@dataclass
class SquadHeadOutput(ModelOutput):
    """
Sylvain Gugger's avatar
Sylvain Gugger committed
1390
    Base class for outputs of question answering models using a :class:`~transformers.modeling_utils.SQuADHead`.
1391
1392
1393

    Args:
        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned if both :obj:`start_positions` and :obj:`end_positions` are provided):
Sylvain Gugger's avatar
Sylvain Gugger committed
1394
1395
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
            losses.
1396
1397
1398
1399
1400
        start_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
        start_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
            Indices for the top config.start_n_top start token possibilities (beam-search).
        end_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
Sylvain Gugger's avatar
Sylvain Gugger committed
1401
1402
            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities
            (beam-search).
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
        end_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
        cls_logits (``torch.FloatTensor`` of shape ``(batch_size,)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
            Log probabilities for the ``is_impossible`` label of the answers.

    """

    loss: Optional[torch.FloatTensor] = None
    start_top_log_probs: Optional[torch.FloatTensor] = None
    start_top_index: Optional[torch.LongTensor] = None
    end_top_log_probs: Optional[torch.FloatTensor] = None
    end_top_index: Optional[torch.LongTensor] = None
    cls_logits: Optional[torch.FloatTensor] = None


thomwolf's avatar
thomwolf committed
1418
class SQuADHead(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
1419
1420
    r"""
    A SQuAD head inspired by XLNet.
1421

Sylvain Gugger's avatar
Sylvain Gugger committed
1422
1423
1424
1425
    Args:
        config (:class:`~transformers.PretrainedConfig`):
            The config used by the model, will be used to grab the :obj:`hidden_size` of the model and the
            :obj:`layer_norm_eps` to use.
thomwolf's avatar
thomwolf committed
1426
    """
1427

thomwolf's avatar
thomwolf committed
1428
    def __init__(self, config):
Julien Chaumond's avatar
Julien Chaumond committed
1429
        super().__init__()
thomwolf's avatar
thomwolf committed
1430
1431
1432
1433
1434
1435
1436
        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)

Sylvain Gugger's avatar
Sylvain Gugger committed
1437
    @replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
1438
    def forward(
1439
        self,
Sylvain Gugger's avatar
Sylvain Gugger committed
1440
1441
1442
1443
1444
1445
        hidden_states: torch.FloatTensor,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        cls_index: Optional[torch.LongTensor] = None,
        is_impossible: Optional[torch.LongTensor] = None,
        p_mask: Optional[torch.FloatTensor] = None,
1446
        return_dict: bool = False,
Sylvain Gugger's avatar
Sylvain Gugger committed
1447
1448
    ) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
        """
Lysandre's avatar
Lysandre committed
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
        Args:
            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
                Final hidden states of the model on the sequence tokens.
            start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                Positions of the first token for the labeled span.
            end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                Positions of the last token for the labeled span.
            cls_index (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                Position of the CLS token for each sentence in the batch. If :obj:`None`, takes the last token.
            is_impossible (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
                Whether the question has a possible answer in the paragraph or not.
            p_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1461
1462
                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
                should be masked.
Lysandre's avatar
Lysandre committed
1463
            return_dict (:obj:`bool`, `optional`, defaults to :obj:`False`):
1464
                Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
Sylvain Gugger's avatar
Sylvain Gugger committed
1465

Lysandre's avatar
Lysandre committed
1466
        Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1467
        """
thomwolf's avatar
thomwolf committed
1468
        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
thomwolf's avatar
thomwolf committed
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491

        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
1492

1493
            return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
thomwolf's avatar
thomwolf committed
1494
1495
1496
1497

        else:
            # during inference, compute the end logits based on beam search
            bsz, slen, hsz = hidden_states.size()
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
            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)
            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)
            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)
thomwolf's avatar
thomwolf committed
1510
1511
            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)
1512
            end_log_probs = F.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
thomwolf's avatar
thomwolf committed
1513

1514
1515
1516
            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)
thomwolf's avatar
thomwolf committed
1517
1518
1519
1520
1521
1522
            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)

1523
            if not return_dict:
1524
1525
1526
1527
1528
1529
1530
1531
1532
                return (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits)
            else:
                return SquadHeadOutput(
                    start_top_log_probs=start_top_log_probs,
                    start_top_index=start_top_index,
                    end_top_log_probs=end_top_log_probs,
                    end_top_index=end_top_index,
                    cls_logits=cls_logits,
                )
thomwolf's avatar
thomwolf committed
1533
1534
1535


class SequenceSummary(nn.Module):
Sylvain Gugger's avatar
Sylvain Gugger committed
1536
1537
1538
1539
1540
    r"""
    Compute a single vector summary of a sequence hidden states.

    Args:
        config (:class:`~transformers.PretrainedConfig`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1541
1542
            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
            config class of your model for the default values it uses):
Sylvain Gugger's avatar
Sylvain Gugger committed
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554

            - **summary_type** (:obj:`str`) -- The method to use to make this summary. Accepted values are:

                - :obj:`"last"` -- Take the last token hidden state (like XLNet)
                - :obj:`"first"` -- Take the first token hidden state (like Bert)
                - :obj:`"mean"` -- Take the mean of all tokens hidden states
                - :obj:`"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
                - :obj:`"attn"` -- Not implemented now, use multi-head attention

            - **summary_use_proj** (:obj:`bool`) -- Add a projection after the vector extraction.
            - **summary_proj_to_labels** (:obj:`bool`) -- If :obj:`True`, the projection outputs to
              :obj:`config.num_labels` classes (otherwise to :obj:`config.hidden_size`).
Sylvain Gugger's avatar
Sylvain Gugger committed
1555
            - **summary_activation** (:obj:`Optional[str]`) -- Set to :obj:`"tanh"` to add a tanh activation to the
Sylvain Gugger's avatar
Sylvain Gugger committed
1556
1557
1558
1559
1560
              output, another string or :obj:`None` will add no activation.
            - **summary_first_dropout** (:obj:`float`) -- Optional dropout probability before the projection and
              activation.
            - **summary_last_dropout** (:obj:`float`)-- Optional dropout probability after the projection and
              activation.
thomwolf's avatar
thomwolf committed
1561
    """
1562

1563
    def __init__(self, config: PretrainedConfig):
Julien Chaumond's avatar
Julien Chaumond committed
1564
        super().__init__()
thomwolf's avatar
thomwolf committed
1565

1566
        self.summary_type = getattr(config, "summary_type", "last")
1567
        if self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
1568
1569
1570
1571
1572
            # 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
1573
        self.summary = Identity()
1574
1575
        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
1576
                num_classes = config.num_labels
thomwolf's avatar
thomwolf committed
1577
1578
1579
1580
            else:
                num_classes = config.hidden_size
            self.summary = nn.Linear(config.hidden_size, num_classes)

1581
        activation_string = getattr(config, "summary_activation", None)
Lysandre's avatar
Lysandre committed
1582
        self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
thomwolf's avatar
thomwolf committed
1583

thomwolf's avatar
thomwolf committed
1584
        self.first_dropout = Identity()
1585
        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
1586
1587
            self.first_dropout = nn.Dropout(config.summary_first_dropout)

thomwolf's avatar
thomwolf committed
1588
        self.last_dropout = Identity()
1589
        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
1590
            self.last_dropout = nn.Dropout(config.summary_last_dropout)
thomwolf's avatar
thomwolf committed
1591

Sylvain Gugger's avatar
Sylvain Gugger committed
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
    def forward(
        self, hidden_states: torch.FloatTensor, cls_index: Optional[torch.LongTensor] = None
    ) -> torch.FloatTensor:
        """
        Compute a single vector summary of a sequence hidden states.

        Args:
            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`[batch_size, seq_len, hidden_size]`):
                The hidden states of the last layer.
            cls_index (:obj:`torch.LongTensor` of shape :obj:`[batch_size]` or :obj:`[batch_size, ...]` where ... are optional leading dimensions of :obj:`hidden_states`, `optional`):
                Used if :obj:`summary_type == "cls_index"` and takes the last token of the sequence as classification
                token.

        Returns:
            :obj:`torch.FloatTensor`: The summary of the sequence hidden states.
thomwolf's avatar
thomwolf committed
1607
        """
1608
        if self.summary_type == "last":
thomwolf's avatar
thomwolf committed
1609
            output = hidden_states[:, -1]
1610
        elif self.summary_type == "first":
thomwolf's avatar
thomwolf committed
1611
            output = hidden_states[:, 0]
1612
        elif self.summary_type == "mean":
thomwolf's avatar
thomwolf committed
1613
            output = hidden_states.mean(dim=1)
1614
        elif self.summary_type == "cls_index":
thomwolf's avatar
thomwolf committed
1615
            if cls_index is None:
Lysandre's avatar
Lysandre committed
1616
1617
1618
1619
1620
                cls_index = torch.full_like(
                    hidden_states[..., :1, :],
                    hidden_states.shape[-2] - 1,
                    dtype=torch.long,
                )
thomwolf's avatar
thomwolf committed
1621
            else:
thomwolf's avatar
thomwolf committed
1622
                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
1623
                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
thomwolf's avatar
thomwolf committed
1624
            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
1625
1626
            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
        elif self.summary_type == "attn":
thomwolf's avatar
thomwolf committed
1627
1628
            raise NotImplementedError

1629
        output = self.first_dropout(output)
thomwolf's avatar
thomwolf committed
1630
1631
        output = self.summary(output)
        output = self.activation(output)
1632
        output = self.last_dropout(output)
thomwolf's avatar
thomwolf committed
1633
1634
1635
1636

        return output


Sylvain Gugger's avatar
Sylvain Gugger committed
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
def prune_linear_layer(layer: torch.nn.Linear, index: torch.LongTensor, dim: int = 0) -> torch.nn.Linear:
    """
    Prune a linear layer to keep only entries in index.

    Used to remove heads.

    Args:
        layer (:obj:`torch.nn.Linear`): The layer to prune.
        index (:obj:`torch.LongTensor`): The indices to keep in the layer.
        dim (:obj:`int`, `optional`, defaults to 0): The dimension on which to keep the indices.

    Returns:
        :obj:`torch.nn.Linear`: The pruned layer as a new layer with :obj:`requires_grad=True`.
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
    """
    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


Sylvain Gugger's avatar
Sylvain Gugger committed
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
def prune_conv1d_layer(layer: Conv1D, index: torch.LongTensor, dim: int = 1) -> Conv1D:
    """
    Prune a Conv1D layer to keep only entries in index. A Conv1D work as a Linear layer (see e.g. BERT) but the weights
    are transposed.

    Used to remove heads.

    Args:
        layer (:class:`~transformers.modeling_utils.Conv1D`): The layer to prune.
        index (:obj:`torch.LongTensor`): The indices to keep in the layer.
        dim (:obj:`int`, `optional`, defaults to 1): The dimension on which to keep the indices.

    Returns:
        :class:`~transformers.modeling_utils.Conv1D`: The pruned layer as a new layer with :obj:`requires_grad=True`.
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
    """
    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
1702
1703


Sylvain Gugger's avatar
Sylvain Gugger committed
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
def prune_layer(
    layer: Union[torch.nn.Linear, Conv1D], index: torch.LongTensor, dim: Optional[int] = None
) -> Union[torch.nn.Linear, Conv1D]:
    """
    Prune a Conv1D or linear layer to keep only entries in index.

    Used to remove heads.

    Args:
        layer (:obj:`Union[torch.nn.Linear, Conv1D]`): The layer to prune.
        index (:obj:`torch.LongTensor`): The indices to keep in the layer.
        dim (:obj:`int`, `optional`): The dimension on which to keep the indices.

    Returns:
Sylvain Gugger's avatar
Sylvain Gugger committed
1718
1719
        :obj:`torch.nn.Linear` or :class:`~transformers.modeling_utils.Conv1D`: The pruned layer as a new layer with
        :obj:`requires_grad=True`.
1720
1721
1722
1723
1724
1725
1726
    """
    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__))
Patrick von Platen's avatar
Patrick von Platen committed
1727
1728
1729


def apply_chunking_to_forward(
1730
    forward_fn: Callable[..., torch.Tensor], chunk_size: int, chunk_dim: int, *input_tensors
Patrick von Platen's avatar
Patrick von Platen committed
1731
1732
) -> torch.Tensor:
    """
1733
1734
1735
1736
1737
    This function chunks the :obj:`input_tensors` into smaller input tensor parts of size :obj:`chunk_size` over the
    dimension :obj:`chunk_dim`. It then applies a layer :obj:`forward_fn` to each chunk independently to save memory.

    If the :obj:`forward_fn` is independent across the :obj:`chunk_dim` this function will yield the same result as
    directly applying :obj:`forward_fn` to :obj:`input_tensors`.
Patrick von Platen's avatar
Patrick von Platen committed
1738
1739

    Args:
1740
1741
        forward_fn (:obj:`Callable[..., torch.Tensor]`):
            The forward function of the model.
1742
1743
1744
1745
1746
        chunk_size (:obj:`int`):
            The chunk size of a chunked tensor: :obj:`num_chunks = len(input_tensors[0]) / chunk_size`.
        chunk_dim (:obj:`int`):
            The dimension over which the :obj:`input_tensors` should be chunked.
        input_tensors (:obj:`Tuple[torch.Tensor]`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1747
1748
            The input tensors of ``forward_fn`` which will be chunked

Patrick von Platen's avatar
Patrick von Platen committed
1749
    Returns:
1750
        :obj:`torch.Tensor`: A tensor with the same shape as the :obj:`forward_fn` would have given if applied`.
Patrick von Platen's avatar
Patrick von Platen committed
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761


    Examples::

        # rename the usual forward() fn to forward_chunk()
        def forward_chunk(self, hidden_states):
            hidden_states = self.decoder(hidden_states)
            return hidden_states

        # implement a chunked forward function
        def forward(self, hidden_states):
1762
            return apply_chunking_to_forward(self.forward_chunk, self.chunk_size_lm_head, self.seq_len_dim, hidden_states)
Patrick von Platen's avatar
Patrick von Platen committed
1763
1764
1765
    """

    assert len(input_tensors) > 0, "{} has to be a tuple/list of tensors".format(input_tensors)
1766
    tensor_shape = input_tensors[0].shape[chunk_dim]
Patrick von Platen's avatar
Patrick von Platen committed
1767
    assert all(
1768
        input_tensor.shape[chunk_dim] == tensor_shape for input_tensor in input_tensors
Patrick von Platen's avatar
Patrick von Platen committed
1769
1770
    ), "All input tenors have to be of the same shape"

1771
    # inspect.signature exist since python 3.5 and is a python method -> no problem with backward compatibility
Patrick von Platen's avatar
Patrick von Platen committed
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
    num_args_in_forward_chunk_fn = len(inspect.signature(forward_fn).parameters)
    assert num_args_in_forward_chunk_fn == len(
        input_tensors
    ), "forward_chunk_fn expects {} arguments, but only {} input tensors are given".format(
        num_args_in_forward_chunk_fn, len(input_tensors)
    )

    if chunk_size > 0:
        assert (
            input_tensors[0].shape[chunk_dim] % chunk_size == 0
        ), "The dimension to be chunked {} has to be a multiple of the chunk size {}".format(
1783
            input_tensors[0].shape[chunk_dim], chunk_size
Patrick von Platen's avatar
Patrick von Platen committed
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
        )

        num_chunks = input_tensors[0].shape[chunk_dim] // chunk_size

        # chunk input tensor into tuples
        input_tensors_chunks = tuple(input_tensor.chunk(num_chunks, dim=chunk_dim) for input_tensor in input_tensors)
        # apply forward fn to every tuple
        output_chunks = tuple(forward_fn(*input_tensors_chunk) for input_tensors_chunk in zip(*input_tensors_chunks))
        # concatenate output at same dimension
        return torch.cat(output_chunks, dim=chunk_dim)

    return forward_fn(*input_tensors)