modeling_electra.py 69.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# 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.
Sylvain Gugger's avatar
Sylvain Gugger committed
15
"""PyTorch ELECTRA model."""
16
17

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
    CausalLMOutputWithCrossAttentions,
40
41
42
43
44
45
    MaskedLMOutput,
    MultipleChoiceModelOutput,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutput,
    TokenClassifierOutput,
)
Sylvain Gugger's avatar
Sylvain Gugger committed
46
from ...modeling_utils import (
47
48
49
50
51
52
    PreTrainedModel,
    SequenceSummary,
    apply_chunking_to_forward,
    find_pruneable_heads_and_indices,
    prune_linear_layer,
)
Sylvain Gugger's avatar
Sylvain Gugger committed
53
54
from ...utils import logging
from .configuration_electra import ElectraConfig
Lysandre Debut's avatar
Lysandre Debut committed
55
56


Lysandre Debut's avatar
Lysandre Debut committed
57
logger = logging.get_logger(__name__)
Lysandre Debut's avatar
Lysandre Debut committed
58

59
_CHECKPOINT_FOR_DOC = "google/electra-small-discriminator"
60
_CONFIG_FOR_DOC = "ElectraConfig"
61
_TOKENIZER_FOR_DOC = "ElectraTokenizer"
Lysandre Debut's avatar
Lysandre Debut committed
62

63
64
65
66
67
68
69
70
71
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
72
73
74


def load_tf_weights_in_electra(model, config, tf_checkpoint_path, discriminator_or_generator="discriminator"):
Lysandre's avatar
Lysandre committed
75
    """Load tf checkpoints in a pytorch model."""
Lysandre Debut's avatar
Lysandre Debut committed
76
77
    try:
        import re
78

Lysandre Debut's avatar
Lysandre Debut committed
79
80
81
82
83
84
85
86
87
        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)
88
    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
Lysandre Debut's avatar
Lysandre Debut committed
89
90
91
92
93
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
94
        logger.info(f"Loading TF weight {name} with shape {shape}")
Lysandre Debut's avatar
Lysandre Debut committed
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
        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):
117
                logger.info(f"Skipping {original_name}")
Lysandre Debut's avatar
Lysandre Debut committed
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
                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:
143
144
                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
145
146
147
            except AssertionError as e:
                e.args += (pointer.shape, array.shape)
                raise
148
            print(f"Initialize PyTorch weight {name}", original_name)
Lysandre Debut's avatar
Lysandre Debut committed
149
150
            pointer.data = torch.from_numpy(array)
        except AttributeError as e:
151
            print(f"Skipping {original_name}", name, e)
Lysandre Debut's avatar
Lysandre Debut committed
152
153
154
155
            continue
    return model


156
class ElectraEmbeddings(nn.Module):
Lysandre Debut's avatar
Lysandre Debut committed
157
158
159
    """Construct the embeddings from word, position and token_type embeddings."""

    def __init__(self, config):
160
        super().__init__()
Lysandre Debut's avatar
Lysandre Debut committed
161
162
163
164
165
166
        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
167
168
169
170
171
        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)))
172
        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
173
174
175
        if version.parse(torch.__version__) > version.parse("1.6.0"):
            self.register_buffer(
                "token_type_ids",
176
                torch.zeros(self.position_ids.size(), dtype=torch.long),
177
178
                persistent=False,
            )
179

Sylvain Gugger's avatar
Sylvain Gugger committed
180
    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
181
182
183
    def forward(
        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
    ):
184
185
186
187
188
189
190
191
        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:
192
            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
193

194
195
196
        # 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
197
        if token_type_ids is None:
198
199
200
201
202
203
            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)
204
205
206
207
208

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

209
210
211
212
        embeddings = inputs_embeds + token_type_embeddings
        if self.position_embedding_type == "absolute":
            position_embeddings = self.position_embeddings(position_ids)
            embeddings += position_embeddings
213
214
215
216
217
        embeddings = self.LayerNorm(embeddings)
        embeddings = self.dropout(embeddings)
        return embeddings


Sylvain Gugger's avatar
Sylvain Gugger committed
218
# Copied from transformers.models.bert.modeling_bert.BertSelfAttention with Bert->Electra
219
class ElectraSelfAttention(nn.Module):
220
    def __init__(self, config, position_embedding_type=None):
221
222
223
        super().__init__()
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
224
225
                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
                f"heads ({config.num_attention_heads})"
226
227
228
229
230
231
232
233
234
235
236
            )

        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)
237
238
239
        self.position_embedding_type = position_embedding_type or getattr(
            config, "position_embedding_type", "absolute"
        )
240
241
242
        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)
243

244
245
        self.is_decoder = config.is_decoder

246
247
248
249
250
251
252
253
254
255
256
257
    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,
258
        past_key_value=None,
259
260
261
262
263
264
265
        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.
266
267
268
269
270
271
        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]
272
            attention_mask = encoder_attention_mask
273
274
275
276
277
278
279
280
281
        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)
282
        else:
283
284
            key_layer = self.transpose_for_scores(self.key(hidden_states))
            value_layer = self.transpose_for_scores(self.value(hidden_states))
285
286

        query_layer = self.transpose_for_scores(mixed_query_layer)
287
288
289
290
291
292
293
294
295
296

        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)
297
298
299

        # 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))
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316

        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

317
318
319
320
321
322
        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.
323
        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339

        # 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,)
340
341
342

        if self.is_decoder:
            outputs = outputs + (past_key_value,)
343
344
345
        return outputs


Sylvain Gugger's avatar
Sylvain Gugger committed
346
# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
347
348
349
350
351
352
353
354
355
356
357
358
359
360
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
361
# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Electra
362
class ElectraAttention(nn.Module):
363
    def __init__(self, config, position_embedding_type=None):
364
        super().__init__()
365
        self.self = ElectraSelfAttention(config, position_embedding_type=position_embedding_type)
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
391
392
393
        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,
394
        past_key_value=None,
395
396
397
398
399
400
401
402
        output_attentions=False,
    ):
        self_outputs = self.self(
            hidden_states,
            attention_mask,
            head_mask,
            encoder_hidden_states,
            encoder_attention_mask,
403
            past_key_value,
404
405
406
407
408
409
410
            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
411
# Copied from transformers.models.bert.modeling_bert.BertIntermediate
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
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
427
# Copied from transformers.models.bert.modeling_bert.BertOutput
428
429
430
431
432
433
434
435
436
437
438
439
440
441
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
442
# Copied from transformers.models.bert.modeling_bert.BertLayer with Bert->Electra
443
444
445
446
447
448
449
450
451
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:
452
453
            if not self.is_decoder:
                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
454
            self.crossattention = ElectraAttention(config, position_embedding_type="absolute")
455
456
457
458
459
460
461
462
463
464
        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,
465
        past_key_value=None,
466
467
        output_attentions=False,
    ):
468
469
        # 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
470
471
472
473
474
        self_attention_outputs = self.attention(
            hidden_states,
            attention_mask,
            head_mask,
            output_attentions=output_attentions,
475
            past_key_value=self_attn_past_key_value,
476
477
478
        )
        attention_output = self_attention_outputs[0]

479
480
481
482
483
484
485
486
        # 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
487
        if self.is_decoder and encoder_hidden_states is not None:
488
489
490
491
            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`"
                )
492
493
494

            # 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
495
496
497
498
499
500
            cross_attention_outputs = self.crossattention(
                attention_output,
                attention_mask,
                head_mask,
                encoder_hidden_states,
                encoder_attention_mask,
501
                cross_attn_past_key_value,
502
503
504
                output_attentions,
            )
            attention_output = cross_attention_outputs[0]
505
506
507
508
509
            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
510
511
512
513
514

        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
515
516
517
518
519

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

520
521
522
523
524
525
526
527
        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
528
# Copied from transformers.models.bert.modeling_bert.BertEncoder with Bert->Electra
529
530
531
532
533
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)])
534
        self.gradient_checkpointing = False
535
536
537
538
539
540
541
542

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
543
544
        past_key_values=None,
        use_cache=None,
545
546
        output_attentions=False,
        output_hidden_states=False,
547
        return_dict=True,
548
549
    ):
        all_hidden_states = () if output_hidden_states else None
550
551
        all_self_attentions = () if output_attentions else None
        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
552
553

        next_decoder_cache = () if use_cache else None
554
555
556
557
558
        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
559
            past_key_value = past_key_values[i] if past_key_values is not None else None
560

561
            if self.gradient_checkpointing and self.training:
562
563

                if use_cache:
564
                    logger.warning(
565
                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
566
567
                    )
                    use_cache = False
568
569
570

                def create_custom_forward(module):
                    def custom_forward(*inputs):
571
                        return module(*inputs, past_key_value, output_attentions)
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589

                    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,
590
                    past_key_value,
591
592
                    output_attentions,
                )
593

594
            hidden_states = layer_outputs[0]
595
596
            if use_cache:
                next_decoder_cache += (layer_outputs[-1],)
597
            if output_attentions:
598
599
600
                all_self_attentions = all_self_attentions + (layer_outputs[1],)
                if self.config.add_cross_attention:
                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
601
602
603
604
605

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
606
607
            return tuple(
                v
608
609
610
611
612
613
614
                for v in [
                    hidden_states,
                    next_decoder_cache,
                    all_hidden_states,
                    all_self_attentions,
                    all_cross_attentions,
                ]
615
616
                if v is not None
            )
617
        return BaseModelOutputWithPastAndCrossAttentions(
618
            last_hidden_state=hidden_states,
619
            past_key_values=next_decoder_cache,
620
621
622
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
            cross_attentions=all_cross_attentions,
623
        )
Lysandre Debut's avatar
Lysandre Debut committed
624
625
626
627
628
629
630
631
632
633
634
635


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

636
    def forward(self, discriminator_hidden_states):
Lysandre Debut's avatar
Lysandre Debut committed
637
638
        hidden_states = self.dense(discriminator_hidden_states)
        hidden_states = get_activation(self.config.hidden_act)(hidden_states)
639
        logits = self.dense_prediction(hidden_states).squeeze(-1)
Lysandre Debut's avatar
Lysandre Debut committed
640
641
642
643
644
645
646
647
648
649

        return logits


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

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

650
        self.LayerNorm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps)
Lysandre Debut's avatar
Lysandre Debut committed
651
652
653
654
655
656
657
658
659
660
        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


661
class ElectraPreTrainedModel(PreTrainedModel):
Sylvain Gugger's avatar
Sylvain Gugger committed
662
663
664
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
Lysandre Debut's avatar
Lysandre Debut committed
665
666
667
668
669
    """

    config_class = ElectraConfig
    load_tf_weights = load_tf_weights_in_electra
    base_model_prefix = "electra"
670
    supports_gradient_checkpointing = True
671
672
    _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
673

Sylvain Gugger's avatar
Sylvain Gugger committed
674
    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
675
    def _init_weights(self, module):
Patrick von Platen's avatar
Patrick von Platen committed
676
        """Initialize the weights"""
677
        if isinstance(module, nn.Linear):
678
679
680
            # 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)
681
682
683
684
685
686
            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_()
687
688
689
690
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

691
692
693
694
    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, ElectraEncoder):
            module.gradient_checkpointing = value

Lysandre Debut's avatar
Lysandre Debut committed
695

696
@dataclass
Sylvain Gugger's avatar
Sylvain Gugger committed
697
class ElectraForPreTrainingOutput(ModelOutput):
698
    """
699
    Output type of [`ElectraForPreTraining`].
700
701

    Args:
702
        loss (*optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
703
            Total loss of the ELECTRA objective.
704
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length)`):
705
            Prediction scores of the head (scores for each token before SoftMax).
706
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Sylvain Gugger's avatar
Sylvain Gugger committed
707
708
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.
709
710

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
711
        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Sylvain Gugger's avatar
Sylvain Gugger committed
712
713
            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
            sequence_length)`.
714
715
716
717
718

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

719
720
    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None
721
722
723
724
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None


Lysandre Debut's avatar
Lysandre Debut committed
725
ELECTRA_START_DOCSTRING = r"""
Sylvain Gugger's avatar
Sylvain Gugger committed
726

Sylvain Gugger's avatar
Sylvain Gugger committed
727
728
729
    This model inherits from [`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
730

Sylvain Gugger's avatar
Sylvain Gugger committed
731
732
733
    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
734
735

    Parameters:
736
        config ([`ElectraConfig`]): Model configuration class with all the parameters of the model.
Sylvain Gugger's avatar
Sylvain Gugger committed
737
            Initializing with a config file does not load the weights associated with the model, only the
Sylvain Gugger's avatar
Sylvain Gugger committed
738
            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
Lysandre Debut's avatar
Lysandre Debut committed
739
740
741
742
"""

ELECTRA_INPUTS_DOCSTRING = r"""
    Args:
743
        input_ids (`torch.LongTensor` of shape `({0})`):
Lysandre Debut's avatar
Lysandre Debut committed
744
745
            Indices of input sequence tokens in the vocabulary.

Sylvain Gugger's avatar
Sylvain Gugger committed
746
747
            Indices can be obtained using [`ElectraTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.
Lysandre Debut's avatar
Lysandre Debut committed
748

749
750
751
            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
Sylvain Gugger's avatar
Sylvain Gugger committed
752
753

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

756
757
            [What are attention masks?](../glossary#attention-mask)
        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
758
759
            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
760

761
762
            - 0 corresponds to a *sentence A* token,
            - 1 corresponds to a *sentence B* token.
Lysandre Debut's avatar
Lysandre Debut committed
763

764
765
            [What are token type IDs?](../glossary#token-type-ids)
        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
766
767
            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
768

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

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

776
        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
777
778
779
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
780
        encoder_hidden_states  (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
781
782
            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.
783
        encoder_attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
784
            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
785
            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
Sylvain Gugger's avatar
Sylvain Gugger committed
786
787
788
789

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

790
791
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
Sylvain Gugger's avatar
Sylvain Gugger committed
792
            tensors for more detail.
793
794
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
Sylvain Gugger's avatar
Sylvain Gugger committed
795
            more detail.
796
797
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
Lysandre Debut's avatar
Lysandre Debut committed
798
799
800
801
802
803
"""


@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 "
804
    "hidden size and embedding size are different. "
Lysandre Debut's avatar
Lysandre Debut committed
805
806
807
808
809
810
811
812
813
814
815
816
    ""
    "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)

817
        self.encoder = ElectraEncoder(config)
Lysandre Debut's avatar
Lysandre Debut committed
818
        self.config = config
819
820
        # Initialize weights and apply final processing
        self.post_init()
Lysandre Debut's avatar
Lysandre Debut committed
821
822
823
824
825
826
827
828

    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
829
830
831
        """
        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
832
833
834
835
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

836
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
837
    @add_code_sample_docstrings(
838
        processor_class=_TOKENIZER_FOR_DOC,
839
        checkpoint=_CHECKPOINT_FOR_DOC,
840
        output_type=BaseModelOutputWithCrossAttentions,
841
842
        config_class=_CONFIG_FOR_DOC,
    )
Lysandre Debut's avatar
Lysandre Debut committed
843
844
845
846
847
848
849
850
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
851
852
853
854
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        past_key_values=None,
        use_cache=None,
855
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
856
        output_hidden_states=None,
857
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
858
    ):
859
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
Joseph Liu's avatar
Joseph Liu committed
860
861
862
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
863
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
864

Lysandre Debut's avatar
Lysandre Debut committed
865
866
867
868
869
870
871
872
873
        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")

874
        batch_size, seq_length = input_shape
Lysandre Debut's avatar
Lysandre Debut committed
875
876
        device = input_ids.device if input_ids is not None else inputs_embeds.device

877
878
879
        # past_key_values_length
        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0

Lysandre Debut's avatar
Lysandre Debut committed
880
881
882
        if attention_mask is None:
            attention_mask = torch.ones(input_shape, device=device)
        if token_type_ids is None:
883
884
885
886
887
888
            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
889
890

        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
891
892
893
894
895
896
897
898
899
900
901
902

        # If a 2D or 3D attention mask is provided for the cross-attention
        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if self.config.is_decoder and encoder_hidden_states is not None:
            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
            if encoder_attention_mask is None:
                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
        else:
            encoder_extended_attention_mask = None

903
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
Lysandre Debut's avatar
Lysandre Debut committed
904
905

        hidden_states = self.embeddings(
906
907
908
909
910
            input_ids=input_ids,
            position_ids=position_ids,
            token_type_ids=token_type_ids,
            inputs_embeds=inputs_embeds,
            past_key_values_length=past_key_values_length,
Lysandre Debut's avatar
Lysandre Debut committed
911
912
913
914
915
        )

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

916
917
918
919
        hidden_states = self.encoder(
            hidden_states,
            attention_mask=extended_attention_mask,
            head_mask=head_mask,
920
921
922
923
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_extended_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
924
            output_attentions=output_attentions,
Joseph Liu's avatar
Joseph Liu committed
925
            output_hidden_states=output_hidden_states,
926
            return_dict=return_dict,
927
        )
Lysandre Debut's avatar
Lysandre Debut committed
928
929
930
931

        return hidden_states


932
933
934
935
936
937
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)
938
939
940
941
        classifier_dropout = (
            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
        )
        self.dropout = nn.Dropout(classifier_dropout)
942
943
944
945
946
947
948
949
950
951
952
953
954
        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
955
956
957
958
    """
    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.
    """,
959
960
961
962
963
964
    ELECTRA_START_DOCSTRING,
)
class ElectraForSequenceClassification(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
965
        self.config = config
966
967
968
        self.electra = ElectraModel(config)
        self.classifier = ElectraClassificationHead(config)

969
970
        # Initialize weights and apply final processing
        self.post_init()
971

972
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
973
    @add_code_sample_docstrings(
974
        processor_class=_TOKENIZER_FOR_DOC,
975
        checkpoint=_CHECKPOINT_FOR_DOC,
976
977
978
        output_type=SequenceClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
979
980
981
982
983
984
985
986
987
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
988
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
989
        output_hidden_states=None,
990
        return_dict=None,
991
992
    ):
        r"""
993
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
994
995
996
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
997
        """
998
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
999

1000
        discriminator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1001
            input_ids,
1002
1003
1004
1005
1006
1007
1008
1009
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
1010
1011
1012
1013
1014
        )

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

1015
        loss = None
1016
        if labels is not None:
1017
1018
1019
1020
1021
1022
1023
1024
1025
            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":
1026
                loss_fct = MSELoss()
1027
1028
1029
1030
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
1031
            elif self.config.problem_type == "single_label_classification":
1032
1033
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1034
1035
1036
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)
1037

1038
        if not return_dict:
1039
1040
1041
1042
1043
1044
1045
1046
1047
            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,
        )
1048
1049


Lysandre Debut's avatar
Lysandre Debut committed
1050
1051
@add_start_docstrings(
    """
1052
    Electra model with a binary classification head on top as used during pretraining for identifying generated tokens.
Lysandre Debut's avatar
Lysandre Debut committed
1053

Sylvain Gugger's avatar
Sylvain Gugger committed
1054
1055
    It is recommended to load the discriminator checkpoint into that model.
    """,
Lysandre Debut's avatar
Lysandre Debut committed
1056
1057
1058
1059
1060
1061
1062
1063
    ELECTRA_START_DOCSTRING,
)
class ElectraForPreTraining(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
        self.discriminator_predictions = ElectraDiscriminatorPredictions(config)
1064
1065
        # Initialize weights and apply final processing
        self.post_init()
Lysandre Debut's avatar
Lysandre Debut committed
1066

1067
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
Sylvain Gugger's avatar
Sylvain Gugger committed
1068
    @replace_return_docstrings(output_type=ElectraForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
Lysandre Debut's avatar
Lysandre Debut committed
1069
1070
1071
1072
1073
1074
1075
1076
1077
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
1078
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1079
        output_hidden_states=None,
1080
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
1081
1082
    ):
        r"""
1083
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1084
1085
            Labels for computing the ELECTRA loss. Input should be a sequence of tokens (see `input_ids` docstring)
            Indices should be in `[0, 1]`:
Sylvain Gugger's avatar
Sylvain Gugger committed
1086
1087
1088

            - 0 indicates the token is an original token,
            - 1 indicates the token was replaced.
Lysandre Debut's avatar
Lysandre Debut committed
1089

Lysandre's avatar
Lysandre committed
1090
        Returns:
Lysandre Debut's avatar
Lysandre Debut committed
1091

1092
        Examples:
Lysandre Debut's avatar
Lysandre Debut committed
1093

1094
1095
1096
        ```python
        >>> from transformers import ElectraTokenizer, ElectraForPreTraining
        >>> import torch
Lysandre Debut's avatar
Lysandre Debut committed
1097

Sylvain Gugger's avatar
Sylvain Gugger committed
1098
1099
        >>> tokenizer = ElectraTokenizer.from_pretrained("google/electra-small-discriminator")
        >>> model = ElectraForPreTraining.from_pretrained("google/electra-small-discriminator")
Lysandre Debut's avatar
Lysandre Debut committed
1100

Sylvain Gugger's avatar
Sylvain Gugger committed
1101
1102
1103
        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(
        ...     0
        >>> )  # Batch size 1
1104
1105
        >>> logits = model(input_ids).logits
        ```"""
1106
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Lysandre Debut's avatar
Lysandre Debut committed
1107
1108

        discriminator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1109
            input_ids,
1110
1111
1112
1113
1114
1115
1116
1117
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
Lysandre Debut's avatar
Lysandre Debut committed
1118
1119
1120
        )
        discriminator_sequence_output = discriminator_hidden_states[0]

1121
        logits = self.discriminator_predictions(discriminator_sequence_output)
Lysandre Debut's avatar
Lysandre Debut committed
1122

1123
        loss = None
Lysandre Debut's avatar
Lysandre Debut committed
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
        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())

1134
        if not return_dict:
1135
1136
            output = (logits,) + discriminator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output
Lysandre Debut's avatar
Lysandre Debut committed
1137

Sylvain Gugger's avatar
Sylvain Gugger committed
1138
        return ElectraForPreTrainingOutput(
1139
1140
1141
1142
1143
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )
Lysandre Debut's avatar
Lysandre Debut committed
1144
1145
1146
1147
1148
1149


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

Sylvain Gugger's avatar
Sylvain Gugger committed
1150
1151
1152
    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
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
    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)
1163
1164
        # Initialize weights and apply final processing
        self.post_init()
Lysandre Debut's avatar
Lysandre Debut committed
1165
1166
1167
1168

    def get_output_embeddings(self):
        return self.generator_lm_head

1169
1170
1171
    def set_output_embeddings(self, word_embeddings):
        self.generator_lm_head = word_embeddings

1172
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1173
    @add_code_sample_docstrings(
1174
        processor_class=_TOKENIZER_FOR_DOC,
1175
        checkpoint=_CHECKPOINT_FOR_DOC,
1176
1177
1178
        output_type=MaskedLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
Lysandre Debut's avatar
Lysandre Debut committed
1179
1180
1181
1182
1183
1184
1185
1186
    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
1187
        labels=None,
1188
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1189
        output_hidden_states=None,
1190
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
1191
1192
    ):
        r"""
1193
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1194
1195
1196
            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
1197
        """
1198
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Lysandre Debut's avatar
Lysandre Debut committed
1199
1200

        generator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1201
            input_ids,
1202
1203
1204
1205
1206
1207
1208
1209
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
Lysandre Debut's avatar
Lysandre Debut committed
1210
1211
1212
1213
1214
1215
        )
        generator_sequence_output = generator_hidden_states[0]

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

1216
        loss = None
Lysandre Debut's avatar
Lysandre Debut committed
1217
        # Masked language modeling softmax layer
Sylvain Gugger's avatar
Sylvain Gugger committed
1218
        if labels is not None:
Lysandre Debut's avatar
Lysandre Debut committed
1219
            loss_fct = nn.CrossEntropyLoss()  # -100 index = padding token
Sylvain Gugger's avatar
Sylvain Gugger committed
1220
            loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
Lysandre Debut's avatar
Lysandre Debut committed
1221

1222
        if not return_dict:
1223
1224
            output = (prediction_scores,) + generator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output
Lysandre Debut's avatar
Lysandre Debut committed
1225

1226
1227
1228
1229
1230
1231
        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
1232
1233
1234
1235
1236
1237


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

Sylvain Gugger's avatar
Sylvain Gugger committed
1238
1239
    Both the discriminator and generator may be loaded into this model.
    """,
Lysandre Debut's avatar
Lysandre Debut committed
1240
1241
1242
1243
1244
1245
1246
    ELECTRA_START_DOCSTRING,
)
class ElectraForTokenClassification(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
1247
1248
1249
1250
        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
1251
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1252
1253
        # Initialize weights and apply final processing
        self.post_init()
Lysandre Debut's avatar
Lysandre Debut committed
1254

1255
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1256
    @add_code_sample_docstrings(
1257
        processor_class=_TOKENIZER_FOR_DOC,
1258
        checkpoint=_CHECKPOINT_FOR_DOC,
1259
1260
1261
        output_type=TokenClassifierOutput,
        config_class=_CONFIG_FOR_DOC,
    )
Lysandre Debut's avatar
Lysandre Debut committed
1262
1263
1264
1265
1266
1267
1268
1269
1270
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
1271
        output_attentions=None,
Joseph Liu's avatar
Joseph Liu committed
1272
        output_hidden_states=None,
1273
        return_dict=None,
Lysandre Debut's avatar
Lysandre Debut committed
1274
1275
    ):
        r"""
1276
1277
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
Lysandre Debut's avatar
Lysandre Debut committed
1278
        """
1279
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Lysandre Debut's avatar
Lysandre Debut committed
1280
1281

        discriminator_hidden_states = self.electra(
Joseph Liu's avatar
Joseph Liu committed
1282
            input_ids,
1283
1284
1285
1286
1287
1288
1289
1290
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
Lysandre Debut's avatar
Lysandre Debut committed
1291
1292
1293
1294
1295
1296
        )
        discriminator_sequence_output = discriminator_hidden_states[0]

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

1297
        loss = None
Lysandre Debut's avatar
Lysandre Debut committed
1298
1299
1300
1301
1302
        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
1303
1304
1305
1306
                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
1307
1308
                loss = loss_fct(active_logits, active_labels)
            else:
LysandreJik's avatar
LysandreJik committed
1309
                loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
Lysandre Debut's avatar
Lysandre Debut committed
1310

1311
        if not return_dict:
1312
1313
            output = (logits,) + discriminator_hidden_states[1:]
            return ((loss,) + output) if loss is not None else output
Lysandre Debut's avatar
Lysandre Debut committed
1314

1315
1316
1317
1318
1319
1320
        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=discriminator_hidden_states.hidden_states,
            attentions=discriminator_hidden_states.attentions,
        )
1321
1322
1323


@add_start_docstrings(
1324
1325
    """
    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
1326
1327
    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
    """,
1328
    ELECTRA_START_DOCSTRING,
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
)
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)

1341
1342
        # Initialize weights and apply final processing
        self.post_init()
1343

1344
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
1345
    @add_code_sample_docstrings(
1346
        processor_class=_TOKENIZER_FOR_DOC,
1347
        checkpoint=_CHECKPOINT_FOR_DOC,
1348
1349
1350
        output_type=QuestionAnsweringModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
    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
1362
        output_hidden_states=None,
1363
        return_dict=None,
1364
1365
    ):
        r"""
1366
        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1367
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
Sylvain Gugger's avatar
Sylvain Gugger committed
1368
1369
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
1370
        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1371
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
Sylvain Gugger's avatar
Sylvain Gugger committed
1372
1373
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
1374
        """
1375
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1376
1377
1378
1379
1380
1381
1382
1383
1384

        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
1385
            output_hidden_states=output_hidden_states,
1386
1387
1388
1389
1390
1391
        )

        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
1392
1393
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()
1394

1395
        total_loss = None
1396
1397
1398
1399
1400
1401
1402
1403
        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)
1404
1405
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)
1406
1407
1408
1409
1410
1411

            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

1412
        if not return_dict:
Lysandre's avatar
Lysandre committed
1413
1414
1415
1416
            output = (
                start_logits,
                end_logits,
            ) + discriminator_hidden_states[1:]
1417
1418
1419
1420
1421
1422
1423
1424
1425
            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
1426
1427
1428


@add_start_docstrings(
Sylvain Gugger's avatar
Sylvain Gugger committed
1429
1430
1431
1432
    """
    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
1433
    ELECTRA_START_DOCSTRING,
Suraj Patil's avatar
Suraj Patil committed
1434
1435
1436
1437
1438
1439
)
class ElectraForMultipleChoice(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.electra = ElectraModel(config)
Julien Plu's avatar
Julien Plu committed
1440
        self.sequence_summary = SequenceSummary(config)
Suraj Patil's avatar
Suraj Patil committed
1441
1442
        self.classifier = nn.Linear(config.hidden_size, 1)

1443
1444
        # Initialize weights and apply final processing
        self.post_init()
Suraj Patil's avatar
Suraj Patil committed
1445

1446
    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
1447
    @add_code_sample_docstrings(
1448
        processor_class=_TOKENIZER_FOR_DOC,
1449
        checkpoint=_CHECKPOINT_FOR_DOC,
1450
1451
1452
        output_type=MultipleChoiceModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
Suraj Patil's avatar
Suraj Patil committed
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
    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,
1463
        output_hidden_states=None,
1464
        return_dict=None,
Suraj Patil's avatar
Suraj Patil committed
1465
1466
    ):
        r"""
1467
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1468
1469
            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
1470
            `input_ids` above)
Suraj Patil's avatar
Suraj Patil committed
1471
        """
1472
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
Suraj Patil's avatar
Suraj Patil committed
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
        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,
1493
            output_hidden_states=output_hidden_states,
1494
            return_dict=return_dict,
Suraj Patil's avatar
Suraj Patil committed
1495
1496
1497
1498
        )

        sequence_output = discriminator_hidden_states[0]

Julien Plu's avatar
Julien Plu committed
1499
        pooled_output = self.sequence_summary(sequence_output)
Suraj Patil's avatar
Suraj Patil committed
1500
1501
1502
        logits = self.classifier(pooled_output)
        reshaped_logits = logits.view(-1, num_choices)

1503
        loss = None
Suraj Patil's avatar
Suraj Patil committed
1504
1505
1506
1507
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(reshaped_logits, labels)

1508
        if not return_dict:
1509
1510
1511
1512
1513
1514
1515
1516
1517
            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,
        )
1518
1519
1520


@add_start_docstrings(
Sylvain Gugger's avatar
Sylvain Gugger committed
1521
    """ELECTRA Model with a `language modeling` head on top for CLM fine-tuning.""", ELECTRA_START_DOCSTRING
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
)
class ElectraForCausalLM(ElectraPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        if not config.is_decoder:
            logger.warning("If you want to use `ElectraLMHeadModel` as a standalone, add `is_decoder=True.`")

        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

    def set_output_embeddings(self, new_embeddings):
        self.generator_lm_head = new_embeddings

    @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        labels=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            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.
        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
            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]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the left-to-right language modeling loss (next word prediction). 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]`
        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.

Sylvain Gugger's avatar
Sylvain Gugger committed
1579
1580
1581
            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
1582
        use_cache (`bool`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1583
1584
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666

        Returns:

        Example:

        ```python
        >>> from transformers import ElectraTokenizer, ElectraForCausalLM, ElectraConfig
        >>> import torch

        >>> tokenizer = ElectraTokenizer.from_pretrained("google/electra-base-generator")
        >>> config = ElectraConfig.from_pretrained("google/electra-base-generator")
        >>> config.is_decoder = True
        >>> model = ElectraForCausalLM.from_pretrained("google/electra-base-generator", config=config)

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

        >>> prediction_logits = outputs.logits
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if labels is not None:
            use_cache = False

        outputs = 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,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]
        prediction_scores = self.generator_lm_head(self.generator_predictions(sequence_output))

        lm_loss = None
        if labels is not None:
            # we are doing next-token prediction; shift prediction scores and input ids by one
            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
            labels = labels[:, 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (prediction_scores,) + outputs[1:]
            return ((lm_loss,) + output) if lm_loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=lm_loss,
            logits=prediction_scores,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            cross_attentions=outputs.cross_attentions,
        )

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.prepare_inputs_for_generation
    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
        input_shape = input_ids.shape
        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
        if attention_mask is None:
            attention_mask = input_ids.new_ones(input_shape)

        # cut decoder_input_ids if past is used
        if past is not None:
            input_ids = input_ids[:, -1:]

        return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}

    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM._reorder_cache
    def _reorder_cache(self, past, beam_idx):
        reordered_past = ()
        for layer_past in past:
            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
        return reordered_past