test_modeling_gpt2.py 14.8 KB
Newer Older
thomwolf's avatar
thomwolf committed
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

thomwolf's avatar
thomwolf committed
16

17
18
import unittest

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():
27
    import torch
28
29
30
    from transformers import (
        GPT2Config,
        GPT2Model,
31
        GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
32
33
34
35
        GPT2LMHeadModel,
        GPT2DoubleHeadsModel,
    )

36

37
38
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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
class GPT2ModelTester:
    def __init__(
        self,
        parent,
        batch_size=14,
        seq_length=7,
        is_training=True,
        use_token_type_ids=True,
        use_input_mask=True,
        use_labels=True,
        use_mc_token_ids=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 = 14
        self.seq_length = 7
        self.is_training = True
        self.use_token_type_ids = True
        self.use_input_mask = True
        self.use_labels = True
        self.use_mc_token_ids = 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
        self.bos_token_id = vocab_size - 1
        self.eos_token_id = vocab_size - 1

    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)

        mc_token_ids = None
        if self.use_mc_token_ids:
            mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)

        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 = GPT2Config(
            vocab_size=self.vocab_size,
            n_embd=self.hidden_size,
            n_layer=self.num_hidden_layers,
            n_head=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,
            n_positions=self.max_position_embeddings,
            n_ctx=self.max_position_embeddings,
            # type_vocab_size=self.type_vocab_size,
Sylvain Gugger's avatar
Sylvain Gugger committed
125
            # initializer_range=self.initializer_range,
126
127
            bos_token_id=self.bos_token_id,
            eos_token_id=self.eos_token_id,
Sylvain Gugger's avatar
Sylvain Gugger committed
128
            return_dict=True,
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
        )

        head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)

        return (
            config,
            input_ids,
            input_mask,
            head_mask,
            token_type_ids,
            mc_token_ids,
            sequence_labels,
            token_labels,
            choice_labels,
        )

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

    def create_and_check_gpt2_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
        model = GPT2Model(config=config)
        model.to(torch_device)
        model.eval()

Sylvain Gugger's avatar
Sylvain Gugger committed
153
154
155
        result = model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
        result = model(input_ids, token_type_ids=token_type_ids)
        result = model(input_ids)
156
157

        self.parent.assertListEqual(
Sylvain Gugger's avatar
Sylvain Gugger committed
158
            list(result["last_hidden_state"].size()), [self.batch_size, self.seq_length, self.hidden_size],
159
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
160
        self.parent.assertEqual(len(result["past_key_values"]), config.n_layer)
161
162
163
164
165
166
167

    def create_and_check_gpt2_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
        model = GPT2Model(config=config)
        model.to(torch_device)
        model.eval()

        # first forward pass
168
169
170
171
172
173
174
        outputs = model(input_ids, token_type_ids=token_type_ids, use_cache=True)
        outputs_use_cache_conf = model(input_ids, token_type_ids=token_type_ids)
        outputs_no_past = model(input_ids, token_type_ids=token_type_ids, use_cache=False)

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

Sylvain Gugger's avatar
Sylvain Gugger committed
175
        output, past = outputs.to_tuple()
176
177
178
179
180
181
182
183
184

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

        # append to next input_ids and token_type_ids
        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
        next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)

Sylvain Gugger's avatar
Sylvain Gugger committed
185
186
        output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
        output_from_past = model(next_tokens, token_type_ids=next_token_types, past=past)["last_hidden_state"]
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208

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

        # test that outputs are equal for slice
        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))

    def create_and_check_gpt2_model_attention_mask_past(
        self, config, input_ids, input_mask, head_mask, token_type_ids, *args
    ):
        model = GPT2Model(config=config)
        model.to(torch_device)
        model.eval()

        # create attention mask
        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
        half_seq_length = self.seq_length // 2
        attn_mask[:, half_seq_length:] = 0

        # first forward pass
Sylvain Gugger's avatar
Sylvain Gugger committed
209
        output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225

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

        # change a random masked slice from input_ids
        random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
        random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
        input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens

        # append to next input_ids and attn_mask
        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
        attn_mask = torch.cat(
            [attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)], dim=1,
        )

        # get two different outputs
Sylvain Gugger's avatar
Sylvain Gugger committed
226
227
        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
        output_from_past = model(next_tokens, past=past, attention_mask=attn_mask)["last_hidden_state"]
228
229
230
231
232
233
234
235
236
237
238
239
240
241

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

        # test that outputs are equal for slice
        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))

    def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
        model = GPT2LMHeadModel(config)
        model.to(torch_device)
        model.eval()

Sylvain Gugger's avatar
Sylvain Gugger committed
242
        result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
243
244
        self.parent.assertListEqual(list(result["loss"].size()), [])
        self.parent.assertListEqual(
Sylvain Gugger's avatar
Sylvain Gugger committed
245
            list(result["logits"].size()), [self.batch_size, self.seq_length, self.vocab_size],
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
        )

    def create_and_check_double_lm_head_model(
        self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args
    ):
        model = GPT2DoubleHeadsModel(config)
        model.to(torch_device)
        model.eval()

        multiple_choice_inputs_ids = input_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()
        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()

        inputs = {
            "input_ids": multiple_choice_inputs_ids,
            "mc_token_ids": mc_token_ids,
            "attention_mask": multiple_choice_input_mask,
            "token_type_ids": multiple_choice_token_type_ids,
            "labels": multiple_choice_inputs_ids,
        }

Sylvain Gugger's avatar
Sylvain Gugger committed
267
268
        result = model(**inputs)
        self.parent.assertListEqual(list(result["lm_loss"].size()), [])
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
        self.parent.assertListEqual(
            list(result["lm_logits"].size()), [self.batch_size, self.num_choices, self.seq_length, self.vocab_size],
        )
        self.parent.assertListEqual(list(result["mc_logits"].size()), [self.batch_size, self.num_choices])

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()

        (
            config,
            input_ids,
            input_mask,
            head_mask,
            token_type_ids,
            mc_token_ids,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = config_and_inputs

        inputs_dict = {
            "input_ids": input_ids,
            "token_type_ids": token_type_ids,
            "head_mask": head_mask,
        }

        return config, inputs_dict


298
@require_torch
299
class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
300

thomwolf's avatar
thomwolf committed
301
    all_model_classes = (GPT2Model, GPT2LMHeadModel, GPT2DoubleHeadsModel) if is_torch_available() else ()
302
303
304
    all_generative_model_classes = (
        (GPT2LMHeadModel,) if is_torch_available() else ()
    )  # TODO (PVP): Add Double HeadsModel when generate() function is changed accordingly
305
    test_missing_keys = False
306
307

    def setUp(self):
308
        self.model_tester = GPT2ModelTester(self)
309
        self.config_tester = ConfigTester(self, config_class=GPT2Config, n_embd=37)
thomwolf's avatar
thomwolf committed
310
311

    def test_config(self):
312
        self.config_tester.run_common_tests()
thomwolf's avatar
thomwolf committed
313

314
315
316
    def test_gpt2_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_gpt2_model(*config_and_inputs)
thomwolf's avatar
thomwolf committed
317

318
319
320
321
322
323
324
325
    def test_gpt2_model_past(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_gpt2_model_past(*config_and_inputs)

    def test_gpt2_model_att_mask_past(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_gpt2_model_attention_mask_past(*config_and_inputs)

326
327
328
329
330
331
332
    def test_gpt2_lm_head_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_lm_head_model(*config_and_inputs)

    def test_gpt2_double_lm_head_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_double_lm_head_model(*config_and_inputs)
thomwolf's avatar
thomwolf committed
333

334
    @slow
335
    def test_model_from_pretrained(self):
336
        for model_name in GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
337
            model = GPT2Model.from_pretrained(model_name)
338
            self.assertIsNotNone(model)
339
340


341
@require_torch
342
343
344
345
class GPT2ModelLanguageGenerationTest(unittest.TestCase):
    @slow
    def test_lm_generate_gpt2(self):
        model = GPT2LMHeadModel.from_pretrained("gpt2")
346
        model.to(torch_device)
patrickvonplaten's avatar
patrickvonplaten committed
347
        input_ids = torch.tensor([[464, 3290]], dtype=torch.long, device=torch_device)  # The dog
348
349
350
        expected_output_ids = [
            464,
            3290,
patrickvonplaten's avatar
patrickvonplaten committed
351
352
353
354
355
356
357
358
359
360
            373,
            1043,
            287,
            257,
            2214,
            1474,
            262,
            16246,
            286,
            2688,
361
            290,
patrickvonplaten's avatar
patrickvonplaten committed
362
363
364
365
366
367
            2688,
            27262,
            13,
            198,
            198,
            464,
368
            3290,
patrickvonplaten's avatar
patrickvonplaten committed
369
370
        ]  # The dog was found in a field near the intersection of West and West Streets.\n\nThe dog
        output_ids = model.generate(input_ids, do_sample=False)
371
372
373
374
375
        self.assertListEqual(output_ids[0].tolist(), expected_output_ids)

    @slow
    def test_lm_generate_distilgpt2(self):
        model = GPT2LMHeadModel.from_pretrained("distilgpt2")
376
        model.to(torch_device)
patrickvonplaten's avatar
patrickvonplaten committed
377
        input_ids = torch.tensor([[464, 1893]], dtype=torch.long, device=torch_device)  # The president
378
379
        expected_output_ids = [
            464,
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
            1893,
            286,
            262,
            1578,
            1829,
            11,
            290,
            262,
            1893,
            286,
            262,
            1578,
            7526,
            11,
            423,
            587,
396
397
            287,
            262,
398
399
            2635,
        ]  # The president of the United States, and the president of the United Kingdom, have been in the White
400

401
        output_ids = model.generate(input_ids, do_sample=False)
402
        self.assertListEqual(output_ids[0].tolist(), expected_output_ids)