test_modeling_tf_bert.py 29.2 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
# coding=utf-8
Sylvain Gugger's avatar
Sylvain Gugger committed
2
# Copyright 2020 The HuggingFace Team. All rights reserved.
thomwolf's avatar
thomwolf committed
3
4
5
6
7
8
9
10
11
12
13
14
#
# 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.
Aymeric Augustin's avatar
Aymeric Augustin committed
15

thomwolf's avatar
thomwolf committed
16

Matt's avatar
Matt committed
17
18
from __future__ import annotations

19
20
import unittest

Aymeric Augustin's avatar
Aymeric Augustin committed
21
from transformers import BertConfig, is_tf_available
22
from transformers.models.auto import get_values
23
from transformers.testing_utils import require_tf, slow
thomwolf's avatar
thomwolf committed
24

Yih-Dar's avatar
Yih-Dar committed
25
26
from ...test_configuration_common import ConfigTester
from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
27
from ...test_pipeline_mixin import PipelineTesterMixin
Yih-Dar's avatar
Yih-Dar committed
28
from ...utils.test_modeling_tf_core import TFCoreModelTesterMixin
thomwolf's avatar
thomwolf committed
29

thomwolf's avatar
thomwolf committed
30

thomwolf's avatar
thomwolf committed
31
if is_tf_available():
thomwolf's avatar
thomwolf committed
32
    import tensorflow as tf
33

34
    from transformers import TF_MODEL_FOR_PRETRAINING_MAPPING
Sylvain Gugger's avatar
Sylvain Gugger committed
35
    from transformers.models.bert.modeling_tf_bert import (
36
        TFBertForMaskedLM,
37
        TFBertForMultipleChoice,
38
39
        TFBertForNextSentencePrediction,
        TFBertForPreTraining,
40
        TFBertForQuestionAnswering,
41
42
        TFBertForSequenceClassification,
        TFBertForTokenClassification,
43
44
        TFBertLMHeadModel,
        TFBertModel,
45
    )
thomwolf's avatar
thomwolf committed
46

thomwolf's avatar
thomwolf committed
47

48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
class TFBertModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        seq_length=7,
        is_training=True,
        use_input_mask=True,
        use_token_type_ids=True,
        use_labels=True,
        vocab_size=99,
        hidden_size=32,
        num_hidden_layers=5,
        num_attention_heads=4,
        intermediate_size=37,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=512,
        type_vocab_size=16,
        type_sequence_label_size=2,
        initializer_range=0.02,
        num_labels=3,
        num_choices=4,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = 13
        self.seq_length = 7
        self.is_training = True
        self.use_input_mask = True
        self.use_token_type_ids = True
        self.use_labels = True
        self.vocab_size = 99
        self.hidden_size = 32
        self.num_hidden_layers = 5
        self.num_attention_heads = 4
        self.intermediate_size = 37
        self.hidden_act = "gelu"
        self.hidden_dropout_prob = 0.1
        self.attention_probs_dropout_prob = 0.1
        self.max_position_embeddings = 512
        self.type_vocab_size = 16
        self.type_sequence_label_size = 2
        self.initializer_range = 0.02
        self.num_labels = 3
        self.num_choices = 4
        self.scope = None
thomwolf's avatar
thomwolf committed
96

97
98
99
100
101
    def prepare_config_and_inputs(self):
        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)

        input_mask = None
        if self.use_input_mask:
102
            input_mask = random_attention_mask([self.batch_size, self.seq_length])
thomwolf's avatar
thomwolf committed
103

104
105
106
        token_type_ids = None
        if self.use_token_type_ids:
            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
thomwolf's avatar
thomwolf committed
107

108
109
110
111
112
113
114
        sequence_labels = None
        token_labels = None
        choice_labels = None
        if self.use_labels:
            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
            choice_labels = ids_tensor([self.batch_size], self.num_choices)
thomwolf's avatar
thomwolf committed
115

116
117
118
119
120
121
122
123
124
125
126
127
128
        config = BertConfig(
            vocab_size=self.vocab_size,
            hidden_size=self.hidden_size,
            num_hidden_layers=self.num_hidden_layers,
            num_attention_heads=self.num_attention_heads,
            intermediate_size=self.intermediate_size,
            hidden_act=self.hidden_act,
            hidden_dropout_prob=self.hidden_dropout_prob,
            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
            max_position_embeddings=self.max_position_embeddings,
            type_vocab_size=self.type_vocab_size,
            initializer_range=self.initializer_range,
        )
thomwolf's avatar
thomwolf committed
129

130
        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
thomwolf's avatar
thomwolf committed
131

132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
    def prepare_config_and_inputs_for_decoder(self):
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = self.prepare_config_and_inputs()

        config.is_decoder = True
        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)

        return (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        )

159
    def create_and_check_model(
160
161
162
163
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = TFBertModel(config=config)
        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
164
        result = model(inputs)
thomwolf's avatar
thomwolf committed
165

166
        inputs = [input_ids, input_mask]
Sylvain Gugger's avatar
Sylvain Gugger committed
167
        result = model(inputs)
thomwolf's avatar
thomwolf committed
168

Sylvain Gugger's avatar
Sylvain Gugger committed
169
        result = model(input_ids)
thomwolf's avatar
thomwolf committed
170

171
172
        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
thomwolf's avatar
thomwolf committed
173

174
    def create_and_check_causal_lm_base_model(
175
176
177
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.is_decoder = True
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228

        model = TFBertModel(config=config)
        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
        result = model(inputs)

        inputs = [input_ids, input_mask]
        result = model(inputs)

        result = model(input_ids)

        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))

    def create_and_check_model_as_decoder(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.add_cross_attention = True

        model = TFBertModel(config=config)
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
            "encoder_hidden_states": encoder_hidden_states,
            "encoder_attention_mask": encoder_attention_mask,
        }
        result = model(inputs)

        inputs = [input_ids, input_mask]
        result = model(inputs, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states)

        # Also check the case where encoder outputs are not passed
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)

        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))

    def create_and_check_causal_lm_model(
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.is_decoder = True

229
230
231
232
233
234
        model = TFBertLMHeadModel(config=config)
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
235
        prediction_scores = model(inputs)["logits"]
236
237
238
239
        self.parent.assertListEqual(
            list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
        )

240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
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
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
    def create_and_check_causal_lm_model_as_decoder(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.add_cross_attention = True

        model = TFBertLMHeadModel(config=config)
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
            "encoder_hidden_states": encoder_hidden_states,
            "encoder_attention_mask": encoder_attention_mask,
        }
        result = model(inputs)

        inputs = [input_ids, input_mask]
        result = model(inputs, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states)

        prediction_scores = result["logits"]
        self.parent.assertListEqual(
            list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
        )

    def create_and_check_causal_lm_model_past(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
    ):
        config.is_decoder = True

        model = TFBertLMHeadModel(config=config)

        # first forward pass
        outputs = model(input_ids, use_cache=True)
        outputs_use_cache_conf = model(input_ids)
        outputs_no_past = model(input_ids, use_cache=False)

        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)

        past_key_values = outputs.past_key_values

        # create hypothetical next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)

        # append to next input_ids and attn_mask
        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)

        output_from_no_past = model(next_input_ids, output_hidden_states=True).hidden_states[0]
        output_from_past = model(
            next_tokens, past_key_values=past_key_values, output_hidden_states=True
        ).hidden_states[0]

        # select random slice
        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
        output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx]
        output_from_past_slice = output_from_past[:, 0, random_slice_idx]

        # test that outputs are equal for slice
        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-6)

    def create_and_check_causal_lm_model_past_with_attn_mask(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
    ):
        config.is_decoder = True

        model = TFBertLMHeadModel(config=config)

        # create attention mask
        half_seq_length = self.seq_length // 2
        attn_mask_begin = tf.ones((self.batch_size, half_seq_length), dtype=tf.int32)
        attn_mask_end = tf.zeros((self.batch_size, self.seq_length - half_seq_length), dtype=tf.int32)
        attn_mask = tf.concat([attn_mask_begin, attn_mask_end], axis=1)

        # first forward pass
        outputs = model(input_ids, attention_mask=attn_mask, use_cache=True)

        # create hypothetical next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)

        past_key_values = outputs.past_key_values

        # change a random masked slice from input_ids
        random_seq_idx_to_change = ids_tensor((1,), half_seq_length).numpy() + 1
        random_other_next_tokens = ids_tensor((self.batch_size, self.seq_length), config.vocab_size)
        vector_condition = tf.range(self.seq_length) == (self.seq_length - random_seq_idx_to_change)
        condition = tf.transpose(
            tf.broadcast_to(tf.expand_dims(vector_condition, -1), (self.seq_length, self.batch_size))
        )
        input_ids = tf.where(condition, random_other_next_tokens, input_ids)

        # append to next input_ids and
        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
        attn_mask = tf.concat(
            [attn_mask, tf.ones((attn_mask.shape[0], 1), dtype=tf.int32)],
            axis=1,
        )

        output_from_no_past = model(
            next_input_ids,
            attention_mask=attn_mask,
            output_hidden_states=True,
        ).hidden_states[0]
        output_from_past = model(
            next_tokens, past_key_values=past_key_values, attention_mask=attn_mask, output_hidden_states=True
        ).hidden_states[0]

        # select random slice
        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
        output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx]
        output_from_past_slice = output_from_past[:, 0, random_slice_idx]

        # test that outputs are equal for slice
        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-6)

    def create_and_check_causal_lm_model_past_large_inputs(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
    ):
        config.is_decoder = True

        model = TFBertLMHeadModel(config=config)

        input_ids = input_ids[:1, :]
        input_mask = input_mask[:1, :]
        self.batch_size = 1

        # first forward pass
        outputs = model(input_ids, attention_mask=input_mask, use_cache=True)
        past_key_values = outputs.past_key_values

        # create hypothetical next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
        next_attn_mask = ids_tensor((self.batch_size, 3), 2)

        # append to next input_ids and
        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
        next_attention_mask = tf.concat([input_mask, next_attn_mask], axis=-1)

        output_from_no_past = model(
            next_input_ids,
            attention_mask=next_attention_mask,
            output_hidden_states=True,
        ).hidden_states[0]
        output_from_past = model(
            next_tokens,
            attention_mask=next_attention_mask,
            past_key_values=past_key_values,
            output_hidden_states=True,
        ).hidden_states[0]

        self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])

        # select random slice
        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
        output_from_past_slice = output_from_past[:, :, random_slice_idx]

        # test that outputs are equal for slice
        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)

    def create_and_check_decoder_model_past_large_inputs(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        config.add_cross_attention = True

        model = TFBertLMHeadModel(config=config)

        input_ids = input_ids[:1, :]
        input_mask = input_mask[:1, :]
        encoder_hidden_states = encoder_hidden_states[:1, :, :]
        encoder_attention_mask = encoder_attention_mask[:1, :]
        self.batch_size = 1

        # first forward pass
        outputs = model(
            input_ids,
            attention_mask=input_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            use_cache=True,
        )
        past_key_values = outputs.past_key_values

        # create hypothetical next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
        next_attn_mask = ids_tensor((self.batch_size, 3), 2)

        # append to next input_ids and
        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
        next_attention_mask = tf.concat([input_mask, next_attn_mask], axis=-1)

        output_from_no_past = model(
            next_input_ids,
            attention_mask=next_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            output_hidden_states=True,
        ).hidden_states[0]
        output_from_past = model(
            next_tokens,
            attention_mask=next_attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            past_key_values=past_key_values,
            output_hidden_states=True,
        ).hidden_states[0]

        self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])

        # select random slice
        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
        output_from_past_slice = output_from_past[:, :, random_slice_idx]

        # test that outputs are equal for slice
        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)

    def create_and_check_for_masked_lm(
495
496
497
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = TFBertForMaskedLM(config=config)
498
499
500
501
502
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
503
        result = model(inputs)
504
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
thomwolf's avatar
thomwolf committed
505

506
    def create_and_check_for_next_sequence_prediction(
507
508
509
510
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = TFBertForNextSentencePrediction(config=config)
        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
Sylvain Gugger's avatar
Sylvain Gugger committed
511
        result = model(inputs)
512
        self.parent.assertEqual(result.logits.shape, (self.batch_size, 2))
thomwolf's avatar
thomwolf committed
513

514
    def create_and_check_for_pretraining(
515
516
517
518
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = TFBertForPreTraining(config=config)
        inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
Sylvain Gugger's avatar
Sylvain Gugger committed
519
        result = model(inputs)
520
521
        self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
        self.parent.assertEqual(result.seq_relationship_logits.shape, (self.batch_size, 2))
thomwolf's avatar
thomwolf committed
522

523
    def create_and_check_for_sequence_classification(
524
525
526
527
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.num_labels = self.num_labels
        model = TFBertForSequenceClassification(config=config)
528
529
530
531
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
532
        }
533

Sylvain Gugger's avatar
Sylvain Gugger committed
534
        result = model(inputs)
535
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
thomwolf's avatar
thomwolf committed
536

537
    def create_and_check_for_multiple_choice(
538
539
540
541
542
543
544
545
546
547
548
549
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.num_choices = self.num_choices
        model = TFBertForMultipleChoice(config=config)
        multiple_choice_inputs_ids = tf.tile(tf.expand_dims(input_ids, 1), (1, self.num_choices, 1))
        multiple_choice_input_mask = tf.tile(tf.expand_dims(input_mask, 1), (1, self.num_choices, 1))
        multiple_choice_token_type_ids = tf.tile(tf.expand_dims(token_type_ids, 1), (1, self.num_choices, 1))
        inputs = {
            "input_ids": multiple_choice_inputs_ids,
            "attention_mask": multiple_choice_input_mask,
            "token_type_ids": multiple_choice_token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
550
        result = model(inputs)
551
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
thomwolf's avatar
thomwolf committed
552

553
    def create_and_check_for_token_classification(
554
555
556
557
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.num_labels = self.num_labels
        model = TFBertForTokenClassification(config=config)
558
559
560
561
562
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
563
        result = model(inputs)
564
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
thomwolf's avatar
thomwolf committed
565

566
    def create_and_check_for_question_answering(
567
568
569
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = TFBertForQuestionAnswering(config=config)
570
571
572
573
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
574
        }
575

Sylvain Gugger's avatar
Sylvain Gugger committed
576
        result = model(inputs)
577
578
        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
thomwolf's avatar
thomwolf committed
579

580
581
582
583
584
585
586
587
588
589
590
591
592
    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = config_and_inputs
        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
        return config, inputs_dict
thomwolf's avatar
thomwolf committed
593
594


595
@require_tf
596
class TFBertModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
597
598
599
600
    all_model_classes = (
        (
            TFBertModel,
            TFBertForMaskedLM,
Lysandre Debut's avatar
Lysandre Debut committed
601
            TFBertLMHeadModel,
602
603
604
605
606
607
608
609
610
611
            TFBertForNextSentencePrediction,
            TFBertForPreTraining,
            TFBertForQuestionAnswering,
            TFBertForSequenceClassification,
            TFBertForTokenClassification,
            TFBertForMultipleChoice,
        )
        if is_tf_available()
        else ()
    )
612
613
614
615
616
617
618
619
620
621
622
623
624
    pipeline_model_mapping = (
        {
            "feature-extraction": TFBertModel,
            "fill-mask": TFBertForMaskedLM,
            "question-answering": TFBertForQuestionAnswering,
            "text-classification": TFBertForSequenceClassification,
            "text-generation": TFBertLMHeadModel,
            "token-classification": TFBertForTokenClassification,
            "zero-shot": TFBertForSequenceClassification,
        }
        if is_tf_available()
        else {}
    )
625
    test_head_masking = False
626
627
    test_onnx = True
    onnx_min_opset = 10
thomwolf's avatar
thomwolf committed
628

629
630
631
632
633
    # special case for ForPreTraining model
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)

        if return_labels:
634
            if model_class in get_values(TF_MODEL_FOR_PRETRAINING_MAPPING):
635
636
637
638
                inputs_dict["next_sentence_label"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)

        return inputs_dict

thomwolf's avatar
thomwolf committed
639
    def setUp(self):
640
        self.model_tester = TFBertModelTester(self)
thomwolf's avatar
thomwolf committed
641
642
643
644
645
        self.config_tester = ConfigTester(self, config_class=BertConfig, hidden_size=37)

    def test_config(self):
        self.config_tester.run_common_tests()

646
647
    def test_model(self):
        """Test the base model"""
thomwolf's avatar
thomwolf committed
648
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
        self.model_tester.create_and_check_model(*config_and_inputs)

    def test_causal_lm_base_model(self):
        """Test the base model of the causal LM model

        is_deocder=True, no cross_attention, no encoder outputs
        """
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_causal_lm_base_model(*config_and_inputs)

    def test_model_as_decoder(self):
        """Test the base model as a decoder (of an encoder-decoder architecture)

        is_deocder=True + cross_attention + pass encoder outputs
        """
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
thomwolf's avatar
thomwolf committed
666
667
668

    def test_for_masked_lm(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
669
        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
thomwolf's avatar
thomwolf committed
670

671
    def test_for_causal_lm(self):
672
        """Test the causal LM model"""
673
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
        self.model_tester.create_and_check_causal_lm_model(*config_and_inputs)

    def test_causal_lm_model_as_decoder(self):
        """Test the causal LM model as a decoder"""
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_causal_lm_model_as_decoder(*config_and_inputs)

    def test_causal_lm_model_past(self):
        """Test causal LM model with `past_key_values`"""
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_causal_lm_model_past(*config_and_inputs)

    def test_causal_lm_model_past_with_attn_mask(self):
        """Test the causal LM model with `past_key_values` and `attention_mask`"""
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_causal_lm_model_past_with_attn_mask(*config_and_inputs)

    def test_causal_lm_model_past_with_large_inputs(self):
        """Test the causal LM model with `past_key_values` and a longer decoder sequence length"""
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_causal_lm_model_past_large_inputs(*config_and_inputs)

    def test_decoder_model_past_with_large_inputs(self):
        """Similar to `test_causal_lm_model_past_with_large_inputs` but with cross-attention"""
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
700

thomwolf's avatar
thomwolf committed
701
702
    def test_for_multiple_choice(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
703
        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
thomwolf's avatar
thomwolf committed
704
705
706

    def test_for_next_sequence_prediction(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
707
        self.model_tester.create_and_check_for_next_sequence_prediction(*config_and_inputs)
thomwolf's avatar
thomwolf committed
708
709
710

    def test_for_pretraining(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
711
        self.model_tester.create_and_check_for_pretraining(*config_and_inputs)
thomwolf's avatar
thomwolf committed
712
713
714

    def test_for_question_answering(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
715
        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
thomwolf's avatar
thomwolf committed
716
717
718

    def test_for_sequence_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
719
        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
thomwolf's avatar
thomwolf committed
720
721
722

    def test_for_token_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
723
        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
thomwolf's avatar
thomwolf committed
724
725

    def test_model_from_pretrained(self):
Julien Plu's avatar
Julien Plu committed
726
727
728
729
730
        model = TFBertModel.from_pretrained("jplu/tiny-tf-bert-random")
        self.assertIsNotNone(model)

    def test_custom_load_tf_weights(self):
        model, output_loading_info = TFBertForTokenClassification.from_pretrained(
Julien Chaumond's avatar
Julien Chaumond committed
731
            "jplu/tiny-tf-bert-random", output_loading_info=True
Julien Plu's avatar
Julien Plu committed
732
        )
Julien Plu's avatar
Julien Plu committed
733
        self.assertEqual(sorted(output_loading_info["unexpected_keys"]), [])
Julien Plu's avatar
Julien Plu committed
734
735
        for layer in output_loading_info["missing_keys"]:
            self.assertTrue(layer.split("_")[0] in ["dropout", "classifier"])
736

737
738
739
740
741
    # TODO (Joao): fix me
    @unittest.skip("Onnx compliancy broke with TF 2.10")
    def test_onnx_compliancy(self):
        pass

742

743
@require_tf
744
745
746
747
748
749
750
class TFBertModelIntegrationTest(unittest.TestCase):
    @slow
    def test_inference_masked_lm(self):
        model = TFBertForPreTraining.from_pretrained("lysandre/tiny-bert-random")
        input_ids = tf.constant([[0, 1, 2, 3, 4, 5]])
        output = model(input_ids)[0]

LysandreJik's avatar
LysandreJik committed
751
        expected_shape = [1, 6, 32000]
752
753
754
755
756
757
758
        self.assertEqual(output.shape, expected_shape)

        print(output[:, :3, :3])

        expected_slice = tf.constant(
            [
                [
LysandreJik's avatar
LysandreJik committed
759
760
761
                    [-0.05243197, -0.04498899, 0.05512108],
                    [-0.07444685, -0.01064632, 0.04352357],
                    [-0.05020351, 0.05530146, 0.00700043],
762
763
764
765
                ]
            ]
        )
        tf.debugging.assert_near(output[:, :3, :3], expected_slice, atol=1e-4)