"vscode:/vscode.git/clone" did not exist on "66e588d837275b26b428f737692357090ad41426"
modeling_electra.py 61.1 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# coding=utf-8
# Copyright 2019 The Google AI Language Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch ELECTRA model. """

import math
Lysandre Debut's avatar
Lysandre Debut committed
18
import os
19
20
from dataclasses import dataclass
from typing import Optional, Tuple
Lysandre Debut's avatar
Lysandre Debut committed
21
22

import torch
23
import torch.utils.checkpoint
24
from packaging import version
25
from torch import nn
26
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
Lysandre Debut's avatar
Lysandre Debut committed
27

Sylvain Gugger's avatar
Sylvain Gugger committed
28
29
from ...activations import ACT2FN, get_activation
from ...file_utils import (
30
31
32
    ModelOutput,
    add_code_sample_docstrings,
    add_start_docstrings,
33
    add_start_docstrings_to_model_forward,
34
35
    replace_return_docstrings,
)
Sylvain Gugger's avatar
Sylvain Gugger committed
36
from ...modeling_outputs import (
37
    BaseModelOutputWithCrossAttentions,
38
    BaseModelOutputWithPastAndCrossAttentions,
39
40
41
42
43
44
    MaskedLMOutput,
    MultipleChoiceModelOutput,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
Sylvain Gugger's avatar
Sylvain Gugger committed
45
from ...modeling_utils import (
46
47
48
49
50
51
    PreTrainedModel,
    SequenceSummary,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
    prune_linear_layer,
)
Sylvain Gugger's avatar
Sylvain Gugger committed
52
53
from ...utils import logging
from .configuration_electra import ElectraConfig
Lysandre Debut's avatar
Lysandre Debut committed
54
55


Lysandre Debut's avatar
Lysandre Debut committed
56
logger = logging.get_logger(__name__)
Lysandre Debut's avatar
Lysandre Debut committed
57

58
_CHECKPOINT_FOR_DOC = "google/electra-small-discriminator"
59
_CONFIG_FOR_DOC = "ElectraConfig"
60
_TOKENIZER_FOR_DOC = "ElectraTokenizer"
Lysandre Debut's avatar
Lysandre Debut committed
61

62
63
64
65
66
67
68
69
70
ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "google/electra-small-generator",
    "google/electra-base-generator",
    "google/electra-large-generator",
    "google/electra-small-discriminator",
    "google/electra-base-discriminator",
    "google/electra-large-discriminator",
    # See all ELECTRA models at https://huggingface.co/models?filter=electra
]
Lysandre Debut's avatar
Lysandre Debut committed
71
72
73


def load_tf_weights_in_electra(model, config, tf_checkpoint_path, discriminator_or_generator="discriminator"):
Lysandre's avatar
Lysandre committed
74
    """Load tf checkpoints in a pytorch model."""
Lysandre Debut's avatar
Lysandre Debut committed
75
76
    try:
        import re
77

Lysandre Debut's avatar
Lysandre Debut committed
78
79
80
81
82
83
84
85
86
        import numpy as np
        import tensorflow as tf
    except ImportError:
        logger.error(
            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions."
        )
        raise
    tf_path = os.path.abspath(tf_checkpoint_path)
87
    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
Lysandre Debut's avatar
Lysandre Debut committed
88
89
90
91
92
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
93
        logger.info(f"Loading TF weight {name} with shape {shape}")
Lysandre Debut's avatar
Lysandre Debut committed
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array)
    for name, array in zip(names, arrays):
        original_name: str = name

        try:
            if isinstance(model, ElectraForMaskedLM):
                name = name.replace("electra/embeddings/", "generator/embeddings/")

            if discriminator_or_generator == "generator":
                name = name.replace("electra/", "discriminator/")
                name = name.replace("generator/", "electra/")

            name = name.replace("dense_1", "dense_prediction")
            name = name.replace("generator_predictions/output_bias", "generator_lm_head/bias")

            name = name.split("/")
            # print(original_name, name)
            # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
            # which are not required for using pretrained model
            if any(n in ["global_step", "temperature"] for n in name):
116
                logger.info(f"Skipping {original_name}")
Lysandre Debut's avatar
Lysandre Debut committed
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
                continue
            pointer = model
            for m_name in name:
                if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
                    scope_names = re.split(r"_(\d+)", m_name)
                else:
                    scope_names = [m_name]
                if scope_names[0] == "kernel" or scope_names[0] == "gamma":
                    pointer = getattr(pointer, "weight")
                elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
                    pointer = getattr(pointer, "bias")
                elif scope_names[0] == "output_weights":
                    pointer = getattr(pointer, "weight")
                elif scope_names[0] == "squad":
                    pointer = getattr(pointer, "classifier")
                else:
                    pointer = getattr(pointer, scope_names[0])
                if len(scope_names) >= 2:
                    num = int(scope_names[1])
                    pointer = pointer[num]
            if m_name.endswith("_embeddings"):
                pointer = getattr(pointer, "weight")
            elif m_name == "kernel":
                array = np.transpose(array)
            try:
142
143
                if pointer.shape != array.shape:
                    raise ValueError(f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched")
Lysandre Debut's avatar
Lysandre Debut committed
144
145
146
            except AssertionError as e:
                e.args += (pointer.shape, array.shape)
                raise
147
            print(f"Initialize PyTorch weight {name}", original_name)
Lysandre Debut's avatar
Lysandre Debut committed
148
149
            pointer.data = torch.from_numpy(array)
        except AttributeError as e:
150
            print(f"Skipping {original_name}", name, e)
Lysandre Debut's avatar
Lysandre Debut committed
151
152
153
154
            continue
    return model


155
class ElectraEmbeddings(nn.Module):
Lysandre Debut's avatar
Lysandre Debut committed
156
157
158
    """Construct the embeddings from word, position and token_type embeddings."""

    def __init__(self, config):
159
        super().__init__()
Lysandre Debut's avatar
Lysandre Debut committed
160
161
162
163
164
165
        self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)

        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
        # any TensorFlow checkpoint file
166
167
168
169
170
        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
171
        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
172
173
174
175
176
177
        if version.parse(torch.__version__) > version.parse("1.6.0"):
            self.register_buffer(
                "token_type_ids",
                torch.zeros(self.position_ids.size(), dtype=torch.long, device=self.position_ids.device),
                persistent=False,
            )
178

Sylvain Gugger's avatar
Sylvain Gugger committed
179
    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
180
181
182
    def forward(
        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
    ):
183
184
185
186
187
188
189
190
        if input_ids is not None:
            input_shape = input_ids.size()
        else:
            input_shape = inputs_embeds.size()[:-1]

        seq_length = input_shape[1]

        if position_ids is None:
191
            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
192

193
194
195
        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
        # issue #5664
196
        if token_type_ids is None:
197
198
199
200
201
202
            if hasattr(self, "token_type_ids"):
                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
                token_type_ids = buffered_token_type_ids_expanded
            else:
                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
203
204
205
206
207

        if inputs_embeds is None:
            inputs_embeds = self.word_embeddings(input_ids)
        token_type_embeddings = self.token_type_embeddings(token_type_ids)

208
209
210
211
        embeddings = inputs_embeds + token_type_embeddings
        if self.position_embedding_type == "absolute":
            position_embeddings = self.position_embeddings(position_ids)
            embeddings += position_embeddings
212
213
214
215
216
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


Sylvain Gugger's avatar
Sylvain Gugger committed
217
# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Electra
218
219
220
221
222
class ElectraSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
223
224
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
225
226
227
228
229
230
231
232
233
234
235
            )

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
236
        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
237
238
239
        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
            self.max_position_embeddings = config.max_position_embeddings
            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
240

241
242
        self.is_decoder = config.is_decoder

243
244
245
246
247
248
249
250
251
252
253
254
    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
255
        past_key_value=None,
256
257
258
259
260
261
262
        output_attentions=False,
    ):
        mixed_query_layer = self.query(hidden_states)

        # If this is instantiated as a cross-attention module, the keys
        # and values come from an encoder; the attention mask needs to be
        # such that the encoder's padding tokens are not attended to.
263
264
265
266
267
268
        is_cross_attention = encoder_hidden_states is not None

        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_layer = past_key_value[0]
            value_layer = past_key_value[1]
269
            attention_mask = encoder_attention_mask
270
271
272
273
274
275
276
277
278
        elif is_cross_attention:
            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
            attention_mask = encoder_attention_mask
        elif past_key_value is not None:
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))
            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
279
        else:
280
281
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))
282
283

        query_layer = self.transpose_for_scores(mixed_query_layer)
284
285
286
287
288
289
290
291
292
293

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_layer, value_layer)
294
295
296

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313

        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
            seq_length = hidden_states.size()[1]
            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
            distance = position_ids_l - position_ids_r
            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility

            if self.position_embedding_type == "relative_key":
                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
                attention_scores = attention_scores + relative_position_scores
            elif self.position_embedding_type == "relative_key_query":
                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key

314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
        if attention_mask is not None:
            # Apply the attention mask is (precomputed for all layers in ElectraModel forward() function)
            attention_scores = attention_scores + attention_mask

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)

        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = attention_probs * head_mask

        context_layer = torch.matmul(attention_probs, value_layer)

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)

        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
337
338
339

        if self.is_decoder:
            outputs = outputs + (past_key_value,)
340
341
342
        return outputs


Sylvain Gugger's avatar
Sylvain Gugger committed
343
# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
344
345
346
347
348
349
350
351
352
353
354
355
356
357
class ElectraSelfOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


Sylvain Gugger's avatar
Sylvain Gugger committed
358
# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Electra
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
class ElectraAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.self = ElectraSelfAttention(config)
        self.output = ElectraSelfOutput(config)
        self.pruned_heads = set()

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
        )

        # Prune linear layers
        self.self.query = prune_linear_layer(self.self.query, index)
        self.self.key = prune_linear_layer(self.self.key, index)
        self.self.value = prune_linear_layer(self.self.value, index)
        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)

        # Update hyper params and store pruned heads
        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
391
        past_key_value=None,
392
393
394
395
396
397
398
399
        output_attentions=False,
    ):
        self_outputs = self.self(
            hidden_states,
            attention_mask,
            head_mask,
            encoder_hidden_states,
            encoder_attention_mask,
400
            past_key_value,
401
402
403
404
405
406
407
            output_attentions,
        )
        attention_output = self.output(self_outputs[0], hidden_states)
        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
        return outputs


Sylvain Gugger's avatar
Sylvain Gugger committed
408
# Copied from transformers.models.bert.modeling_bert.BertIntermediate
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
class ElectraIntermediate(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = ACT2FN[config.hidden_act]
        else:
            self.intermediate_act_fn = config.hidden_act

    def forward(self, hidden_states):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)
        return hidden_states


Sylvain Gugger's avatar
Sylvain Gugger committed
424
# Copied from transformers.models.bert.modeling_bert.BertOutput
425
426
427
428
429
430
431
432
433
434
435
436
437
438
class ElectraOutput(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.LayerNorm(hidden_states + input_tensor)
        return hidden_states


Sylvain Gugger's avatar
Sylvain Gugger committed
439
# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Electra
440
441
442
443
444
445
446
447
448
class ElectraLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = ElectraAttention(config)
        self.is_decoder = config.is_decoder
        self.add_cross_attention = config.add_cross_attention
        if self.add_cross_attention:
449
450
            if not self.is_decoder:
                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
451
452
453
454
455
456
457
458
459
460
461
            self.crossattention = ElectraAttention(config)
        self.intermediate = ElectraIntermediate(config)
        self.output = ElectraOutput(config)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
462
        past_key_value=None,
463
464
        output_attentions=False,
    ):
465
466
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
467
468
469
470
471
        self_attention_outputs = self.attention(
            hidden_states,
            attention_mask,
            head_mask,
            output_attentions=output_attentions,
472
            past_key_value=self_attn_past_key_value,
473
474
475
        )
        attention_output = self_attention_outputs[0]

476
477
478
479
480
481
482
483
        # if decoder, the last output is tuple of self-attn cache
        if self.is_decoder:
            outputs = self_attention_outputs[1:-1]
            present_key_value = self_attention_outputs[-1]
        else:
            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights

        cross_attn_present_key_value = None
484
        if self.is_decoder and encoder_hidden_states is not None:
485
486
487
488
            if not hasattr(self, "crossattention"):
                raise ValueError(
                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
                )
489
490
491

            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
492
493
494
495
496
497
            cross_attention_outputs = self.crossattention(
                attention_output,
                attention_mask,
                head_mask,
                encoder_hidden_states,
                encoder_attention_mask,
498
                cross_attn_past_key_value,
499
500
501
                output_attentions,
            )
            attention_output = cross_attention_outputs[0]
502
503
504
505
506
            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights

            # add cross-attn cache to positions 3,4 of present_key_value tuple
            cross_attn_present_key_value = cross_attention_outputs[-1]
            present_key_value = present_key_value + cross_attn_present_key_value
507
508
509
510
511

        layer_output = apply_chunking_to_forward(
            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
        )
        outputs = (layer_output,) + outputs
512
513
514
515
516

        # if decoder, return the attn key/values as the last output
        if self.is_decoder:
            outputs = outputs + (present_key_value,)

517
518
519
520
521
522
523
524
        return outputs

    def feed_forward_chunk(self, attention_output):
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output, attention_output)
        return layer_output


Sylvain Gugger's avatar
Sylvain Gugger committed
525
# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Electra
526
527
528
529
530
class ElectraEncoder(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.layer = nn.ModuleList([ElectraLayer(config) for _ in range(config.num_hidden_layers)])
531
        self.gradient_checkpointing = False
532
533
534
535
536
537
538
539

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
540
541
        past_key_values=None,
        use_cache=None,
542
543
        output_attentions=False,
        output_hidden_states=False,
544
        return_dict=True,
545
546
    ):
        all_hidden_states = () if output_hidden_states else None
547
548
        all_self_attentions = () if output_attentions else None
        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
549
550

        next_decoder_cache = () if use_cache else None
551
552
553
554
555
        for i, layer_module in enumerate(self.layer):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            layer_head_mask = head_mask[i] if head_mask is not None else None
556
            past_key_value = past_key_values[i] if past_key_values is not None else None
557

558
            if self.gradient_checkpointing and self.training:
559
560

                if use_cache:
561
                    logger.warning(
562
                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
563
564
                    )
                    use_cache = False
565
566
567

                def create_custom_forward(module):
                    def custom_forward(*inputs):
568
                        return module(*inputs, past_key_value, output_attentions)
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586

                    return custom_forward

                layer_outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(layer_module),
                    hidden_states,
                    attention_mask,
                    layer_head_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                )
            else:
                layer_outputs = layer_module(
                    hidden_states,
                    attention_mask,
                    layer_head_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
587
                    past_key_value,
588
589
                    output_attentions,
                )
590

591
            hidden_states = layer_outputs[0]
592
593
            if use_cache:
                next_decoder_cache += (layer_outputs[-1],)
594
            if output_attentions:
595
596
597
                all_self_attentions = all_self_attentions + (layer_outputs[1],)
                if self.config.add_cross_attention:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
598
599
600
601
602

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
603
604
            return tuple(
                v
605
606
607
608
609
610
611
                for v in [
                    hidden_states,
                    next_decoder_cache,
                    all_hidden_states,
                    all_self_attentions,
                    all_cross_attentions,
                ]
612
613
                if v is not None
            )
614
        return BaseModelOutputWithPastAndCrossAttentions(
615
            last_hidden_state=hidden_states,
616
            past_key_values=next_decoder_cache,
617
618
619
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
            cross_attentions=all_cross_attentions,
620
        )
Lysandre Debut's avatar
Lysandre Debut committed
621
622
623
624
625
626
627
628
629
630
631
632


class ElectraDiscriminatorPredictions(nn.Module):
    """Prediction module for the discriminator, made up of two dense layers."""

    def __init__(self, config):
        super().__init__()

        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dense_prediction = nn.Linear(config.hidden_size, 1)
        self.config = config

633
    def forward(self, discriminator_hidden_states):
Lysandre Debut's avatar
Lysandre Debut committed
634
635
        hidden_states = self.dense(discriminator_hidden_states)
        hidden_states = get_activation(self.config.hidden_act)(hidden_states)
636
        logits = self.dense_prediction(hidden_states).squeeze(-1)
Lysandre Debut's avatar
Lysandre Debut committed
637
638
639
640
641
642
643
644
645
646

        return logits


class ElectraGeneratorPredictions(nn.Module):
    """Prediction module for the generator, made up of two dense layers."""

    def __init__(self, config):
        super().__init__()

647
        self.LayerNorm = nn.LayerNorm(config.embedding_size)
Lysandre Debut's avatar
Lysandre Debut committed
648
649
650
651
652
653
654
655
656
657
        self.dense = nn.Linear(config.hidden_size, config.embedding_size)

    def forward(self, generator_hidden_states):
        hidden_states = self.dense(generator_hidden_states)
        hidden_states = get_activation("gelu")(hidden_states)
        hidden_states = self.LayerNorm(hidden_states)

        return hidden_states


658
class ElectraPreTrainedModel(PreTrainedModel):
Sylvain Gugger's avatar
Sylvain Gugger committed
659
660
661
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
Lysandre Debut's avatar
Lysandre Debut committed
662
663
664
665
666
    """

    config_class = ElectraConfig
    load_tf_weights = load_tf_weights_in_electra
    base_model_prefix = "electra"
667
    supports_gradient_checkpointing = True
668
669
    _keys_to_ignore_on_load_missing = [r"position_ids"]
    _keys_to_ignore_on_load_unexpected = [r"electra\.embeddings_project\.weight", r"electra\.embeddings_project\.bias"]
Lysandre Debut's avatar
Lysandre Debut committed
670

Sylvain Gugger's avatar
Sylvain Gugger committed
671
    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
672
    def _init_weights(self, module):
Patrick von Platen's avatar
Patrick von Platen committed
673
        """Initialize the weights"""
674
        if isinstance(module, nn.Linear):
675
676
677
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
678
679
680
681
682
683
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
684
685
686
687
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

688
689
690
691
    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, ElectraEncoder):
            module.gradient_checkpointing = value

Lysandre Debut's avatar
Lysandre Debut committed
692

693
@dataclass
Sylvain Gugger's avatar
Sylvain Gugger committed
694
class ElectraForPreTrainingOutput(ModelOutput):
695
    """
696
    Output type of :class:`~transformers.ElectraForPreTraining`.
697
698

    Args:
699
        loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
700
            Total loss of the ELECTRA objective.
701
        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
702
703
704
705
706
707
708
            Prediction scores of the head (scores for each token before SoftMax).
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
Sylvain Gugger's avatar
Sylvain Gugger committed
709
710
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
            sequence_length, sequence_length)`.
711
712
713
714
715

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
    """

716
717
    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
718
719
720
721
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


Lysandre Debut's avatar
Lysandre Debut committed
722
ELECTRA_START_DOCSTRING = r"""
Sylvain Gugger's avatar
Sylvain Gugger committed
723
724
725
726
727

    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
    pruning heads etc.)

Sylvain Gugger's avatar
Sylvain Gugger committed
728
729
730
    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
    general usage and behavior.
Lysandre Debut's avatar
Lysandre Debut committed
731
732
733

    Parameters:
        config (:class:`~transformers.ElectraConfig`): Model configuration class with all the parameters of the model.
Sylvain Gugger's avatar
Sylvain Gugger committed
734
735
736
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
            weights.
Lysandre Debut's avatar
Lysandre Debut committed
737
738
739
740
"""

ELECTRA_INPUTS_DOCSTRING = r"""
    Args:
Sylvain Gugger's avatar
Sylvain Gugger committed
741
        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
Lysandre Debut's avatar
Lysandre Debut committed
742
743
            Indices of input sequence tokens in the vocabulary.

Sylvain Gugger's avatar
Sylvain Gugger committed
744
745
746
            Indices can be obtained using :class:`~transformers.ElectraTokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
            details.
Lysandre Debut's avatar
Lysandre Debut committed
747
748

            `What are input IDs? <../glossary.html#input-ids>`__
Sylvain Gugger's avatar
Sylvain Gugger committed
749
        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
750
            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
Sylvain Gugger's avatar
Sylvain Gugger committed
751
752

            - 1 for tokens that are **not masked**,
753
            - 0 for tokens that are **masked**.
Lysandre Debut's avatar
Lysandre Debut committed
754
755

            `What are attention masks? <../glossary.html#attention-mask>`__
Sylvain Gugger's avatar
Sylvain Gugger committed
756
        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
757
758
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
            1]``:
Sylvain Gugger's avatar
Sylvain Gugger committed
759
760
761

            - 0 corresponds to a `sentence A` token,
            - 1 corresponds to a `sentence B` token.
Lysandre Debut's avatar
Lysandre Debut committed
762
763

            `What are token type IDs? <../glossary.html#token-type-ids>`_
Sylvain Gugger's avatar
Sylvain Gugger committed
764
        position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
765
766
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
            config.max_position_embeddings - 1]``.
Lysandre Debut's avatar
Lysandre Debut committed
767
768

            `What are position IDs? <../glossary.html#position-ids>`_
769
        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
770
            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
Sylvain Gugger's avatar
Sylvain Gugger committed
771
772
773
774
775

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
Lysandre Debut's avatar
Lysandre Debut committed
776
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
Sylvain Gugger's avatar
Sylvain Gugger committed
777
778
779
            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
            vectors than the model's internal embedding lookup matrix.
        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
780
781
            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
            the model is configured as a decoder.
Sylvain Gugger's avatar
Sylvain Gugger committed
782
        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
783
784
            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
Sylvain Gugger's avatar
Sylvain Gugger committed
785
786
787
788

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

789
        output_attentions (:obj:`bool`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
790
791
            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
            tensors for more detail.
792
        output_hidden_states (:obj:`bool`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
793
794
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
795
        return_dict (:obj:`bool`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
796
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
Lysandre Debut's avatar
Lysandre Debut committed
797
798
799
800
801
802
"""


@add_start_docstrings(
    "The bare Electra Model transformer outputting raw hidden-states without any specific head on top. Identical to "
    "the BERT model except that it uses an additional linear layer between the embedding layer and the encoder if the "
803
    "hidden size and embedding size are different. "
Lysandre Debut's avatar
Lysandre Debut committed
804
805
806
807
808
809
810
811
812
813
814
815
    ""
    "Both the generator and discriminator checkpoints may be loaded into this model.",
    ELECTRA_START_DOCSTRING,
)
class ElectraModel(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.embeddings = ElectraEmbeddings(config)

        if config.embedding_size != config.hidden_size:
            self.embeddings_project = nn.Linear(config.embedding_size, config.hidden_size)

816
        self.encoder = ElectraEncoder(config)
Lysandre Debut's avatar
Lysandre Debut committed
817
818
819
820
821
822
823
824
825
826
        self.config = config
        self.init_weights()

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    def _prune_heads(self, heads_to_prune):
Sylvain Gugger's avatar
Sylvain Gugger committed
827
828
829
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
Lysandre Debut's avatar
Lysandre Debut committed
830
831
832
833
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

834
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
835
    @add_code_sample_docstrings(
836
        processor_class=_TOKENIZER_FOR_DOC,
837
        checkpoint=_CHECKPOINT_FOR_DOC,
838
        output_type=BaseModelOutputWithCrossAttentions,
839
840
        config_class=_CONFIG_FOR_DOC,
    )
Lysandre Debut's avatar
Lysandre Debut committed
841
842
843
844
845
846
847
848
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
849
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
850
        output_hidden_states=None,
851
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
852
    ):
853
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
Joseph Liu's avatar
Joseph Liu committed
854
855
856
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
857
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
858

Lysandre Debut's avatar
Lysandre Debut committed
859
860
861
862
863
864
865
866
867
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

868
        batch_size, seq_length = input_shape
Lysandre Debut's avatar
Lysandre Debut committed
869
870
871
872
873
        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if attention_mask is None:
            attention_mask = torch.ones(input_shape, device=device)
        if token_type_ids is None:
874
875
876
877
878
879
            if hasattr(self.embeddings, "token_type_ids"):
                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
                token_type_ids = buffered_token_type_ids_expanded
            else:
                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
Lysandre Debut's avatar
Lysandre Debut committed
880
881

        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
882
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
Lysandre Debut's avatar
Lysandre Debut committed
883
884
885
886
887
888
889
890

        hidden_states = self.embeddings(
            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
        )

        if hasattr(self, "embeddings_project"):
            hidden_states = self.embeddings_project(hidden_states)

891
892
893
894
895
        hidden_states = self.encoder(
            hidden_states,
            attention_mask=extended_attention_mask,
            head_mask=head_mask,
            output_attentions=output_attentions,
Joseph Liu's avatar
Joseph Liu committed
896
            output_hidden_states=output_hidden_states,
897
            return_dict=return_dict,
898
        )
Lysandre Debut's avatar
Lysandre Debut committed
899
900
901
902

        return hidden_states


903
904
905
906
907
908
class ElectraClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
909
910
911
912
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
913
914
915
916
917
918
919
920
921
922
923
924
925
        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)

    def forward(self, features, **kwargs):
        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
        x = self.dropout(x)
        x = self.dense(x)
        x = get_activation("gelu")(x)  # although BERT uses tanh here, it seems Electra authors used gelu here
        x = self.dropout(x)
        x = self.out_proj(x)
        return x


@add_start_docstrings(
Sylvain Gugger's avatar
Sylvain Gugger committed
926
927
928
929
    """
    ELECTRA Model transformer with a sequence classification/regression head on top (a linear layer on top of the
    pooled output) e.g. for GLUE tasks.
    """,
930
931
932
933
934
935
    ELECTRA_START_DOCSTRING,
)
class ElectraForSequenceClassification(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
936
        self.config = config
937
938
939
940
941
        self.electra = ElectraModel(config)
        self.classifier = ElectraClassificationHead(config)

        self.init_weights()

942
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
943
    @add_code_sample_docstrings(
944
        processor_class=_TOKENIZER_FOR_DOC,
945
        checkpoint=_CHECKPOINT_FOR_DOC,
946
947
948
        output_type=SequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
949
950
951
952
953
954
955
956
957
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
958
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
959
        output_hidden_states=None,
960
        return_dict=None,
961
962
    ):
        r"""
963
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
964
965
            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
966
967
            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
968
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
969

970
        discriminator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
971
972
973
974
975
976
977
978
            input_ids,
            attention_mask,
            token_type_ids,
            position_ids,
            head_mask,
            inputs_embeds,
            output_attentions,
            output_hidden_states,
979
            return_dict,
980
981
982
983
984
        )

        sequence_output = discriminator_hidden_states[0]
        logits = self.classifier(sequence_output)

985
        loss = None
986
        if labels is not None:
987
988
989
990
991
992
993
994
995
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
996
                loss_fct = MSELoss()
997
998
999
1000
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
1001
            elif self.config.problem_type == "single_label_classification":
1002
1003
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1004
1005
1006
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)
1007

1008
        if not return_dict:
1009
1010
1011
1012
1013
1014
1015
1016
1017
            output = (logits,) + discriminator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )
1018
1019


Lysandre Debut's avatar
Lysandre Debut committed
1020
1021
@add_start_docstrings(
    """
1022
    Electra model with a binary classification head on top as used during pretraining for identifying generated tokens.
Lysandre Debut's avatar
Lysandre Debut committed
1023

Sylvain Gugger's avatar
Sylvain Gugger committed
1024
1025
    It is recommended to load the discriminator checkpoint into that model.
    """,
Lysandre Debut's avatar
Lysandre Debut committed
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
    ELECTRA_START_DOCSTRING,
)
class ElectraForPreTraining(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
        self.discriminator_predictions = ElectraDiscriminatorPredictions(config)
        self.init_weights()

1036
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
Sylvain Gugger's avatar
Sylvain Gugger committed
1037
    @replace_return_docstrings(output_type=ElectraForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
Lysandre Debut's avatar
Lysandre Debut committed
1038
1039
1040
1041
1042
1043
1044
1045
1046
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
1047
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1048
        output_hidden_states=None,
1049
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
1050
1051
    ):
        r"""
Sylvain Gugger's avatar
Sylvain Gugger committed
1052
        labels (``torch.LongTensor`` of shape ``(batch_size, sequence_length)``, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1053
1054
            Labels for computing the ELECTRA loss. Input should be a sequence of tokens (see :obj:`input_ids`
            docstring) Indices should be in ``[0, 1]``:
Sylvain Gugger's avatar
Sylvain Gugger committed
1055
1056
1057

            - 0 indicates the token is an original token,
            - 1 indicates the token was replaced.
Lysandre Debut's avatar
Lysandre Debut committed
1058

Lysandre's avatar
Lysandre committed
1059
        Returns:
Lysandre Debut's avatar
Lysandre Debut committed
1060

Lysandre's avatar
Lysandre committed
1061
        Examples::
Lysandre Debut's avatar
Lysandre Debut committed
1062

Lysandre's avatar
Lysandre committed
1063
1064
            >>> from transformers import ElectraTokenizer, ElectraForPreTraining
            >>> import torch
Lysandre Debut's avatar
Lysandre Debut committed
1065

Lysandre's avatar
Lysandre committed
1066
1067
            >>> tokenizer = ElectraTokenizer.from_pretrained('google/electra-small-discriminator')
            >>> model = ElectraForPreTraining.from_pretrained('google/electra-small-discriminator')
Lysandre Debut's avatar
Lysandre Debut committed
1068

Lysandre's avatar
Lysandre committed
1069
1070
            >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
            >>> logits = model(input_ids).logits
Lysandre Debut's avatar
Lysandre Debut committed
1071
        """
1072
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Lysandre Debut's avatar
Lysandre Debut committed
1073
1074

        discriminator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1075
1076
1077
1078
1079
1080
1081
1082
            input_ids,
            attention_mask,
            token_type_ids,
            position_ids,
            head_mask,
            inputs_embeds,
            output_attentions,
            output_hidden_states,
1083
            return_dict,
Lysandre Debut's avatar
Lysandre Debut committed
1084
1085
1086
        )
        discriminator_sequence_output = discriminator_hidden_states[0]

1087
        logits = self.discriminator_predictions(discriminator_sequence_output)
Lysandre Debut's avatar
Lysandre Debut committed
1088

1089
        loss = None
Lysandre Debut's avatar
Lysandre Debut committed
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
        if labels is not None:
            loss_fct = nn.BCEWithLogitsLoss()
            if attention_mask is not None:
                active_loss = attention_mask.view(-1, discriminator_sequence_output.shape[1]) == 1
                active_logits = logits.view(-1, discriminator_sequence_output.shape[1])[active_loss]
                active_labels = labels[active_loss]
                loss = loss_fct(active_logits, active_labels.float())
            else:
                loss = loss_fct(logits.view(-1, discriminator_sequence_output.shape[1]), labels.float())

1100
        if not return_dict:
1101
1102
            output = (logits,) + discriminator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output
Lysandre Debut's avatar
Lysandre Debut committed
1103

Sylvain Gugger's avatar
Sylvain Gugger committed
1104
        return ElectraForPreTrainingOutput(
1105
1106
1107
1108
1109
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )
Lysandre Debut's avatar
Lysandre Debut committed
1110
1111
1112
1113
1114
1115


@add_start_docstrings(
    """
    Electra model with a language modeling head on top.

Sylvain Gugger's avatar
Sylvain Gugger committed
1116
1117
1118
    Even though both the discriminator and generator may be loaded into this model, the generator is the only model of
    the two to have been trained for the masked language modeling task.
    """,
Lysandre Debut's avatar
Lysandre Debut committed
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
    ELECTRA_START_DOCSTRING,
)
class ElectraForMaskedLM(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
        self.generator_predictions = ElectraGeneratorPredictions(config)

        self.generator_lm_head = nn.Linear(config.embedding_size, config.vocab_size)
        self.init_weights()

    def get_output_embeddings(self):
        return self.generator_lm_head

1134
1135
1136
    def set_output_embeddings(self, word_embeddings):
        self.generator_lm_head = word_embeddings

1137
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1138
    @add_code_sample_docstrings(
1139
        processor_class=_TOKENIZER_FOR_DOC,
1140
        checkpoint=_CHECKPOINT_FOR_DOC,
1141
1142
1143
        output_type=MaskedLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
Lysandre Debut's avatar
Lysandre Debut committed
1144
1145
1146
1147
1148
1149
1150
1151
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
Sylvain Gugger's avatar
Sylvain Gugger committed
1152
        labels=None,
1153
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1154
        output_hidden_states=None,
1155
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
1156
1157
    ):
        r"""
1158
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1159
1160
1161
            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
Lysandre Debut's avatar
Lysandre Debut committed
1162
        """
1163
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Lysandre Debut's avatar
Lysandre Debut committed
1164
1165

        generator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1166
1167
1168
1169
1170
1171
1172
1173
            input_ids,
            attention_mask,
            token_type_ids,
            position_ids,
            head_mask,
            inputs_embeds,
            output_attentions,
            output_hidden_states,
1174
            return_dict,
Lysandre Debut's avatar
Lysandre Debut committed
1175
1176
1177
1178
1179
1180
        )
        generator_sequence_output = generator_hidden_states[0]

        prediction_scores = self.generator_predictions(generator_sequence_output)
        prediction_scores = self.generator_lm_head(prediction_scores)

1181
        loss = None
Lysandre Debut's avatar
Lysandre Debut committed
1182
        # Masked language modeling softmax layer
Sylvain Gugger's avatar
Sylvain Gugger committed
1183
        if labels is not None:
Lysandre Debut's avatar
Lysandre Debut committed
1184
            loss_fct = nn.CrossEntropyLoss()  # -100 index = padding token
Sylvain Gugger's avatar
Sylvain Gugger committed
1185
            loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
Lysandre Debut's avatar
Lysandre Debut committed
1186

1187
        if not return_dict:
1188
1189
            output = (prediction_scores,) + generator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output
Lysandre Debut's avatar
Lysandre Debut committed
1190

1191
1192
1193
1194
1195
1196
        return MaskedLMOutput(
            loss=loss,
            logits=prediction_scores,
            hidden_states=generator_hidden_states.hidden_states,
            attentions=generator_hidden_states.attentions,
        )
Lysandre Debut's avatar
Lysandre Debut committed
1197
1198
1199
1200
1201
1202


@add_start_docstrings(
    """
    Electra model with a token classification head on top.

Sylvain Gugger's avatar
Sylvain Gugger committed
1203
1204
    Both the discriminator and generator may be loaded into this model.
    """,
Lysandre Debut's avatar
Lysandre Debut committed
1205
1206
1207
1208
1209
1210
1211
    ELECTRA_START_DOCSTRING,
)
class ElectraForTokenClassification(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
1212
1213
1214
1215
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
Lysandre Debut's avatar
Lysandre Debut committed
1216
1217
1218
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
        self.init_weights()

1219
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1220
    @add_code_sample_docstrings(
1221
        processor_class=_TOKENIZER_FOR_DOC,
1222
        checkpoint=_CHECKPOINT_FOR_DOC,
1223
1224
1225
        output_type=TokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
Lysandre Debut's avatar
Lysandre Debut committed
1226
1227
1228
1229
1230
1231
1232
1233
1234
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
1235
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1236
        output_hidden_states=None,
1237
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
1238
1239
    ):
        r"""
1240
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1241
1242
            Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -
            1]``.
Lysandre Debut's avatar
Lysandre Debut committed
1243
        """
1244
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Lysandre Debut's avatar
Lysandre Debut committed
1245
1246

        discriminator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1247
1248
1249
1250
1251
1252
1253
1254
            input_ids,
            attention_mask,
            token_type_ids,
            position_ids,
            head_mask,
            inputs_embeds,
            output_attentions,
            output_hidden_states,
1255
            return_dict,
Lysandre Debut's avatar
Lysandre Debut committed
1256
1257
1258
1259
1260
1261
        )
        discriminator_sequence_output = discriminator_hidden_states[0]

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

1262
        loss = None
Lysandre Debut's avatar
Lysandre Debut committed
1263
1264
1265
1266
1267
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            # Only keep active parts of the loss
            if attention_mask is not None:
                active_loss = attention_mask.view(-1) == 1
1268
1269
1270
1271
                active_logits = logits.view(-1, self.config.num_labels)
                active_labels = torch.where(
                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
                )
Lysandre Debut's avatar
Lysandre Debut committed
1272
1273
                loss = loss_fct(active_logits, active_labels)
            else:
LysandreJik's avatar
LysandreJik committed
1274
                loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
Lysandre Debut's avatar
Lysandre Debut committed
1275

1276
        if not return_dict:
1277
1278
            output = (logits,) + discriminator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output
Lysandre Debut's avatar
Lysandre Debut committed
1279

1280
1281
1282
1283
1284
1285
        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )
1286
1287
1288


@add_start_docstrings(
1289
1290
    """
    ELECTRA Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
Sylvain Gugger's avatar
Sylvain Gugger committed
1291
1292
    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
    """,
1293
    ELECTRA_START_DOCSTRING,
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
)
class ElectraForQuestionAnswering(ElectraPreTrainedModel):
    config_class = ElectraConfig
    base_model_prefix = "electra"

    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.electra = ElectraModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        self.init_weights()

1308
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1309
    @add_code_sample_docstrings(
1310
        processor_class=_TOKENIZER_FOR_DOC,
1311
        checkpoint=_CHECKPOINT_FOR_DOC,
1312
1313
1314
        output_type=QuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        start_positions=None,
        end_positions=None,
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1326
        output_hidden_states=None,
1327
        return_dict=None,
1328
1329
    ):
        r"""
1330
        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
1331
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
Sylvain Gugger's avatar
Sylvain Gugger committed
1332
1333
            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
            sequence are not taken into account for computing the loss.
1334
        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
1335
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
Sylvain Gugger's avatar
Sylvain Gugger committed
1336
1337
            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
            sequence are not taken into account for computing the loss.
1338
        """
1339
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1340
1341
1342
1343
1344
1345
1346
1347
1348

        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
Joseph Liu's avatar
Joseph Liu committed
1349
            output_hidden_states=output_hidden_states,
1350
1351
1352
1353
1354
1355
        )

        sequence_output = discriminator_hidden_states[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
Fan Zhang's avatar
Fan Zhang committed
1356
1357
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()
1358

1359
        total_loss = None
1360
1361
1362
1363
1364
1365
1366
1367
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
1368
1369
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)
1370
1371
1372
1373
1374
1375

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

1376
        if not return_dict:
Lysandre's avatar
Lysandre committed
1377
1378
1379
1380
            output = (
                start_logits,
                end_logits,
            ) + discriminator_hidden_states[1:]
1381
1382
1383
1384
1385
1386
1387
1388
1389
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )
Suraj Patil's avatar
Suraj Patil committed
1390
1391
1392


@add_start_docstrings(
Sylvain Gugger's avatar
Sylvain Gugger committed
1393
1394
1395
1396
    """
    ELECTRA Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
    softmax) e.g. for RocStories/SWAG tasks.
    """,
Sylvain Gugger's avatar
Sylvain Gugger committed
1397
    ELECTRA_START_DOCSTRING,
Suraj Patil's avatar
Suraj Patil committed
1398
1399
1400
1401
1402
1403
)
class ElectraForMultipleChoice(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
Julien Plu's avatar
Julien Plu committed
1404
        self.sequence_summary = SequenceSummary(config)
Suraj Patil's avatar
Suraj Patil committed
1405
1406
1407
1408
        self.classifier = nn.Linear(config.hidden_size, 1)

        self.init_weights()

1409
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
1410
    @add_code_sample_docstrings(
1411
        processor_class=_TOKENIZER_FOR_DOC,
1412
        checkpoint=_CHECKPOINT_FOR_DOC,
1413
1414
1415
        output_type=MultipleChoiceModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
Suraj Patil's avatar
Suraj Patil committed
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        output_attentions=None,
1426
        output_hidden_states=None,
1427
        return_dict=None,
Suraj Patil's avatar
Suraj Patil committed
1428
1429
    ):
        r"""
1430
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1431
1432
1433
            Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
            num_choices-1]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
            :obj:`input_ids` above)
Suraj Patil's avatar
Suraj Patil committed
1434
        """
1435
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Suraj Patil's avatar
Suraj Patil committed
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]

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

        discriminator_hidden_states = self.electra(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
1456
            output_hidden_states=output_hidden_states,
1457
            return_dict=return_dict,
Suraj Patil's avatar
Suraj Patil committed
1458
1459
1460
1461
        )

        sequence_output = discriminator_hidden_states[0]

Julien Plu's avatar
Julien Plu committed
1462
        pooled_output = self.sequence_summary(sequence_output)
Suraj Patil's avatar
Suraj Patil committed
1463
1464
1465
        logits = self.classifier(pooled_output)
        reshaped_logits = logits.view(-1, num_choices)

1466
        loss = None
Suraj Patil's avatar
Suraj Patil committed
1467
1468
1469
1470
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(reshaped_logits, labels)

1471
        if not return_dict:
1472
1473
1474
1475
1476
1477
1478
1479
1480
            output = (reshaped_logits,) + discriminator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output

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