test_modeling_roberta.py 17.6 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors.
#
# 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

16
17

import unittest
18

19
from transformers import is_torch_available
20
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
thomwolf's avatar
thomwolf committed
21

22
from .test_configuration_common import ConfigTester
23
from .test_generation_utils import GenerationTesterMixin
24
from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
Aymeric Augustin's avatar
Aymeric Augustin committed
25
26


27
if is_torch_available():
thomwolf's avatar
thomwolf committed
28
    import torch
29

30
31
    from transformers import (
        RobertaConfig,
32
        RobertaForCausalLM,
33
        RobertaForMaskedLM,
34
35
        RobertaForMultipleChoice,
        RobertaForQuestionAnswering,
36
37
        RobertaForSequenceClassification,
        RobertaForTokenClassification,
38
39
40
41
42
43
        RobertaModel,
    )
    from transformers.modeling_roberta import (
        ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
        RobertaEmbeddings,
        create_position_ids_from_input_ids,
44
    )
45
46


47
48
class RobertaModelTester:
    def __init__(
Lysandre's avatar
Lysandre committed
49
50
        self,
        parent,
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
    ):
        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

    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:
80
            input_mask = random_attention_mask([self.batch_size, self.seq_length])
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105

        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)

        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)

        config = RobertaConfig(
            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,
Sylvain Gugger's avatar
Sylvain Gugger committed
106
            return_dict=True,
107
108
109
110
        )

        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels

111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
    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,
        )

    def create_and_check_model(
139
140
141
142
143
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = RobertaModel(config=config)
        model.to(torch_device)
        model.eval()
Sylvain Gugger's avatar
Sylvain Gugger committed
144
145
146
147
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
        result = model(input_ids, token_type_ids=token_type_ids)
        result = model(input_ids)

Stas Bekman's avatar
Stas Bekman committed
148
149
        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))
150

151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
    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 = RobertaModel(config)
        model.to(torch_device)
        model.eval()
        result = model(
            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(
            input_ids,
            attention_mask=input_mask,
            token_type_ids=token_type_ids,
            encoder_hidden_states=encoder_hidden_states,
        )
        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_for_causal_lm(
        self,
        config,
        input_ids,
        token_type_ids,
        input_mask,
        sequence_labels,
        token_labels,
        choice_labels,
        encoder_hidden_states,
        encoder_attention_mask,
    ):
        model = RobertaForCausalLM(config=config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))

    def create_and_check_for_masked_lm(
203
204
205
206
207
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = RobertaForMaskedLM(config=config)
        model.to(torch_device)
        model.eval()
Sylvain Gugger's avatar
Sylvain Gugger committed
208
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
Stas Bekman's avatar
Stas Bekman committed
209
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
210

211
    def create_and_check_for_token_classification(
212
213
214
215
216
217
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.num_labels = self.num_labels
        model = RobertaForTokenClassification(config=config)
        model.to(torch_device)
        model.eval()
Sylvain Gugger's avatar
Sylvain Gugger committed
218
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
Stas Bekman's avatar
Stas Bekman committed
219
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
220

221
    def create_and_check_for_multiple_choice(
222
223
224
225
226
227
228
229
230
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        config.num_choices = self.num_choices
        model = RobertaForMultipleChoice(config=config)
        model.to(torch_device)
        model.eval()
        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
Sylvain Gugger's avatar
Sylvain Gugger committed
231
        result = model(
232
233
234
235
236
            multiple_choice_inputs_ids,
            attention_mask=multiple_choice_input_mask,
            token_type_ids=multiple_choice_token_type_ids,
            labels=choice_labels,
        )
Stas Bekman's avatar
Stas Bekman committed
237
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
238

239
    def create_and_check_for_question_answering(
240
241
242
243
244
        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
    ):
        model = RobertaForQuestionAnswering(config=config)
        model.to(torch_device)
        model.eval()
Sylvain Gugger's avatar
Sylvain Gugger committed
245
        result = model(
246
247
248
249
250
251
            input_ids,
            attention_mask=input_mask,
            token_type_ids=token_type_ids,
            start_positions=sequence_labels,
            end_positions=sequence_labels,
        )
Stas Bekman's avatar
Stas Bekman committed
252
253
        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))
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269

    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


270
@require_torch
271
class RobertaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
272

273
274
    all_model_classes = (
        (
275
            RobertaForCausalLM,
276
277
278
279
280
281
282
283
284
285
            RobertaForMaskedLM,
            RobertaModel,
            RobertaForSequenceClassification,
            RobertaForTokenClassification,
            RobertaForMultipleChoice,
            RobertaForQuestionAnswering,
        )
        if is_torch_available()
        else ()
    )
286
    all_generative_model_classes = (RobertaForCausalLM,) if is_torch_available() else ()
287
288

    def setUp(self):
289
        self.model_tester = RobertaModelTester(self)
290
291
292
293
294
        self.config_tester = ConfigTester(self, config_class=RobertaConfig, hidden_size=37)

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

295
    def test_model(self):
296
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
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
        self.model_tester.create_and_check_model(*config_and_inputs)

    def test_model_as_decoder(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)

    def test_model_as_decoder_with_default_input_mask(self):
        # This regression test was failing with PyTorch < 1.3
        (
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        ) = self.model_tester.prepare_config_and_inputs_for_decoder()

        input_mask = None

        self.model_tester.create_and_check_model_as_decoder(
            config,
            input_ids,
            token_type_ids,
            input_mask,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        )

    def test_for_causal_lm(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
        self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)
334
335
336

    def test_for_masked_lm(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
337
        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
338

Lysandre's avatar
Lysandre committed
339
340
    def test_for_token_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
341
        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
Lysandre's avatar
Lysandre committed
342
343
344

    def test_for_multiple_choice(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
345
        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
Lysandre's avatar
Lysandre committed
346
347
348

    def test_for_question_answering(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
349
        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
Lysandre's avatar
Lysandre committed
350

351
    @slow
352
    def test_model_from_pretrained(self):
353
        for model_name in ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
354
            model = RobertaModel.from_pretrained(model_name)
355
356
            self.assertIsNotNone(model)

Dom Hudson's avatar
Dom Hudson committed
357
    def test_create_position_ids_respects_padding_index(self):
Lysandre's avatar
Lysandre committed
358
        """Ensure that the default position ids only assign a sequential . This is a regression
Dom Hudson's avatar
Dom Hudson committed
359
360
361
362
363
364
365
366
367
        test for https://github.com/huggingface/transformers/issues/1761

        The position ids should be masked with the embedding object's padding index. Therefore, the
        first available non-padding position index is RobertaEmbeddings.padding_idx + 1
        """
        config = self.model_tester.prepare_config_and_inputs()[0]
        model = RobertaEmbeddings(config=config)

        input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
368
369
370
        expected_positions = torch.as_tensor(
            [[0 + model.padding_idx + 1, 1 + model.padding_idx + 1, 2 + model.padding_idx + 1, model.padding_idx]]
        )
Dom Hudson's avatar
Dom Hudson committed
371

Sam Shleifer's avatar
Sam Shleifer committed
372
        position_ids = create_position_ids_from_input_ids(input_ids, model.padding_idx)
373
        self.assertEqual(position_ids.shape, expected_positions.shape)
Dom Hudson's avatar
Dom Hudson committed
374
375
376
        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))

    def test_create_position_ids_from_inputs_embeds(self):
Lysandre's avatar
Lysandre committed
377
        """Ensure that the default position ids only assign a sequential . This is a regression
Dom Hudson's avatar
Dom Hudson committed
378
379
380
381
382
383
        test for https://github.com/huggingface/transformers/issues/1761

        The position ids should be masked with the embedding object's padding index. Therefore, the
        first available non-padding position index is RobertaEmbeddings.padding_idx + 1
        """
        config = self.model_tester.prepare_config_and_inputs()[0]
384
385
386
387
388
389
390
391
392
393
394
        embeddings = RobertaEmbeddings(config=config)

        inputs_embeds = torch.Tensor(2, 4, 30)
        expected_single_positions = [
            0 + embeddings.padding_idx + 1,
            1 + embeddings.padding_idx + 1,
            2 + embeddings.padding_idx + 1,
            3 + embeddings.padding_idx + 1,
        ]
        expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
        position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
395
396
        self.assertEqual(position_ids.shape, expected_positions.shape)
        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
397
398


399
400
@require_sentencepiece
@require_tokenizers
Lysandre Debut's avatar
Lysandre Debut committed
401
@require_torch
402
class RobertaModelIntegrationTest(unittest.TestCase):
403
    @slow
404
    def test_inference_masked_lm(self):
405
        model = RobertaForMaskedLM.from_pretrained("roberta-base")
406

407
        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
408
409
        output = model(input_ids)[0]
        expected_shape = torch.Size((1, 11, 50265))
410
        self.assertEqual(output.shape, expected_shape)
411
        # compare the actual values for a slice.
412
413
        expected_slice = torch.tensor(
            [[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
414
        )
415
416
417
418
419
420

        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.base')
        # roberta.eval()
        # expected_slice = roberta.model.forward(input_ids)[0][:, :3, :3].detach()

        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
421

422
    @slow
423
    def test_inference_no_head(self):
424
        model = RobertaModel.from_pretrained("roberta-base")
425

426
        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
427
428
        output = model(input_ids)[0]
        # compare the actual values for a slice.
429
430
        expected_slice = torch.tensor(
            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
431
        )
432
433
434
435
436
437

        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.base')
        # roberta.eval()
        # expected_slice = roberta.extract_features(input_ids)[:, :3, :3].detach()

        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
438

439
    @slow
440
    def test_inference_classification_head(self):
441
        model = RobertaForSequenceClassification.from_pretrained("roberta-large-mnli")
442

443
        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
444
445
        output = model(input_ids)[0]
        expected_shape = torch.Size((1, 3))
446
        self.assertEqual(output.shape, expected_shape)
447
448
449
450
451
452
453
        expected_tensor = torch.tensor([[-0.9469, 0.3913, 0.5118]])

        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.large.mnli')
        # roberta.eval()
        # expected_tensor = roberta.predict("mnli", input_ids, return_logits=True).detach()

        self.assertTrue(torch.allclose(output, expected_tensor, atol=1e-4))