test_modeling_roberta.py 14 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_torch, slow, torch_device
thomwolf's avatar
thomwolf committed
21

22
from .test_configuration_common import ConfigTester
23
from .test_modeling_common import ModelTesterMixin, ids_tensor
Aymeric Augustin's avatar
Aymeric Augustin committed
24
25


26
if is_torch_available():
thomwolf's avatar
thomwolf committed
27
    import torch
28
29
30
31
    from transformers import (
        RobertaConfig,
        RobertaModel,
        RobertaForMaskedLM,
32
33
        RobertaForMultipleChoice,
        RobertaForQuestionAnswering,
34
35
36
        RobertaForSequenceClassification,
        RobertaForTokenClassification,
    )
37
    from transformers.modeling_roberta import RobertaEmbeddings, create_position_ids_from_input_ids
38
    from transformers.modeling_roberta import ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST
39
40


41
42
43
44
45
46
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
96
97
98
class RobertaModelTester:
    def __init__(
        self, parent,
    ):
        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:
            input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)

        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
99
            return_dict=True,
100
101
102
103
104
105
106
107
108
109
110
111
112
        )

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

    def check_loss_output(self, result):
        self.parent.assertListEqual(list(result["loss"].size()), [])

    def create_and_check_roberta_model(
        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
113
114
115
116
        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)

117
        self.parent.assertListEqual(
Sylvain Gugger's avatar
Sylvain Gugger committed
118
            list(result["last_hidden_state"].size()), [self.batch_size, self.seq_length, self.hidden_size]
119
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
120
        self.parent.assertListEqual(list(result["pooler_output"].size()), [self.batch_size, self.hidden_size])
121
122
123
124
125
126
127

    def create_and_check_roberta_for_masked_lm(
        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
128
129
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
        self.parent.assertListEqual(list(result["logits"].size()), [self.batch_size, self.seq_length, self.vocab_size])
130
131
132
133
134
135
136
137
138
        self.check_loss_output(result)

    def create_and_check_roberta_for_token_classification(
        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
139
        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
140
141
142
143
144
145
146
147
148
149
150
151
152
        self.parent.assertListEqual(list(result["logits"].size()), [self.batch_size, self.seq_length, self.num_labels])
        self.check_loss_output(result)

    def create_and_check_roberta_for_multiple_choice(
        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
153
        result = model(
154
155
156
157
158
159
160
161
162
163
164
165
166
167
            multiple_choice_inputs_ids,
            attention_mask=multiple_choice_input_mask,
            token_type_ids=multiple_choice_token_type_ids,
            labels=choice_labels,
        )
        self.parent.assertListEqual(list(result["logits"].size()), [self.batch_size, self.num_choices])
        self.check_loss_output(result)

    def create_and_check_roberta_for_question_answering(
        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
168
        result = model(
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
            input_ids,
            attention_mask=input_mask,
            token_type_ids=token_type_ids,
            start_positions=sequence_labels,
            end_positions=sequence_labels,
        )
        self.parent.assertListEqual(list(result["start_logits"].size()), [self.batch_size, self.seq_length])
        self.parent.assertListEqual(list(result["end_logits"].size()), [self.batch_size, self.seq_length])
        self.check_loss_output(result)

    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


194
@require_torch
195
class RobertaModelTest(ModelTesterMixin, unittest.TestCase):
196

197
198
199
200
201
202
203
204
205
206
207
208
    all_model_classes = (
        (
            RobertaForMaskedLM,
            RobertaModel,
            RobertaForSequenceClassification,
            RobertaForTokenClassification,
            RobertaForMultipleChoice,
            RobertaForQuestionAnswering,
        )
        if is_torch_available()
        else ()
    )
209
210

    def setUp(self):
211
        self.model_tester = RobertaModelTester(self)
212
213
214
215
216
217
218
219
220
221
222
223
224
        self.config_tester = ConfigTester(self, config_class=RobertaConfig, hidden_size=37)

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

    def test_roberta_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_roberta_model(*config_and_inputs)

    def test_for_masked_lm(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_roberta_for_masked_lm(*config_and_inputs)

Lysandre's avatar
Lysandre committed
225
226
227
228
229
230
231
232
233
234
235
236
    def test_for_token_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_roberta_for_token_classification(*config_and_inputs)

    def test_for_multiple_choice(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_roberta_for_multiple_choice(*config_and_inputs)

    def test_for_question_answering(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_roberta_for_question_answering(*config_and_inputs)

237
    @slow
238
    def test_model_from_pretrained(self):
239
        for model_name in ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
240
            model = RobertaModel.from_pretrained(model_name)
241
242
            self.assertIsNotNone(model)

Dom Hudson's avatar
Dom Hudson committed
243
244
245
246
247
248
249
250
251
252
253
    def test_create_position_ids_respects_padding_index(self):
        """ Ensure that the default position ids only assign a sequential . This is a regression
        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]])
254
255
256
        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
257

Sam Shleifer's avatar
Sam Shleifer committed
258
        position_ids = create_position_ids_from_input_ids(input_ids, model.padding_idx)
259
        self.assertEqual(position_ids.shape, expected_positions.shape)
Dom Hudson's avatar
Dom Hudson committed
260
261
262
263
264
265
266
267
268
269
        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))

    def test_create_position_ids_from_inputs_embeds(self):
        """ Ensure that the default position ids only assign a sequential . This is a regression
        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]
270
271
272
273
274
275
276
277
278
279
280
        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)
281
282
        self.assertEqual(position_ids.shape, expected_positions.shape)
        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
283
284
285


class RobertaModelIntegrationTest(unittest.TestCase):
286
    @slow
287
    def test_inference_masked_lm(self):
288
        model = RobertaForMaskedLM.from_pretrained("roberta-base")
289

290
        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
291
292
        output = model(input_ids)[0]
        expected_shape = torch.Size((1, 11, 50265))
293
        self.assertEqual(output.shape, expected_shape)
294
        # compare the actual values for a slice.
295
296
        expected_slice = torch.tensor(
            [[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
297
        )
298
299
300
301
302
303

        # 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))
304

305
    @slow
306
    def test_inference_no_head(self):
307
        model = RobertaModel.from_pretrained("roberta-base")
308

309
        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
310
311
        output = model(input_ids)[0]
        # compare the actual values for a slice.
312
313
        expected_slice = torch.tensor(
            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
314
        )
315
316
317
318
319
320

        # 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))
321

322
    @slow
323
    def test_inference_classification_head(self):
324
        model = RobertaForSequenceClassification.from_pretrained("roberta-large-mnli")
325

326
        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
327
328
        output = model(input_ids)[0]
        expected_shape = torch.Size((1, 3))
329
        self.assertEqual(output.shape, expected_shape)
330
331
332
333
334
335
336
        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))