test_modeling_tf_gpt2.py 13.9 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

Aymeric Augustin's avatar
Aymeric Augustin committed
19
from transformers import GPT2Config, is_tf_available
20
from transformers.testing_utils import require_tf, slow
thomwolf's avatar
thomwolf committed
21

22
from .test_configuration_common import ConfigTester
23
from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
thomwolf's avatar
thomwolf committed
24
25


26
if is_tf_available():
thomwolf's avatar
thomwolf committed
27
    import tensorflow as tf
28
29
30
31
    from transformers.modeling_tf_gpt2 import (
        TFGPT2Model,
        TFGPT2LMHeadModel,
        TFGPT2DoubleHeadsModel,
32
        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
33
        shape_list,
34
    )
thomwolf's avatar
thomwolf committed
35
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
class TFGPT2ModelTester:
    def __init__(
        self, parent,
    ):
        self.parent = parent
        self.batch_size = 13
        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 = self.vocab_size - 1
        self.eos_token_id = self.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,
            # initializer_range=self.initializer_range
            bos_token_id=self.bos_token_id,
            eos_token_id=self.eos_token_id,
Sylvain Gugger's avatar
Sylvain Gugger committed
105
            return_dict=True,
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
        )

        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 create_and_check_gpt2_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
        model = TFGPT2Model(config=config)
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
129
        result = model(inputs)
130
131

        inputs = [input_ids, None, input_mask]  # None is the input for 'past'
Sylvain Gugger's avatar
Sylvain Gugger committed
132
        result = model(inputs)
133

Sylvain Gugger's avatar
Sylvain Gugger committed
134
        result = model(input_ids)
135
136

        self.parent.assertListEqual(
Sylvain Gugger's avatar
Sylvain Gugger committed
137
            list(result["last_hidden_state"].shape), [self.batch_size, self.seq_length, self.hidden_size],
138
139
140
141
142
143
        )

    def create_and_check_gpt2_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
        model = TFGPT2Model(config=config)

        # first forward pass
144
145
146
147
148
149
150
        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
151
        output, past = outputs.to_tuple()
152
153
154
155
156
157
158
159
160

        # 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 = tf.concat([input_ids, next_tokens], axis=-1)
        next_token_type_ids = tf.concat([token_type_ids, next_token_types], axis=-1)

Sylvain Gugger's avatar
Sylvain Gugger committed
161
162
        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"]
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183

        # select random slice
        random_slice_idx = int(ids_tensor((1,), shape_list(output_from_past)[-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_gpt2_model_attention_mask_past(
        self, config, input_ids, input_mask, head_mask, token_type_ids, *args
    ):
        model = TFGPT2Model(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
Sylvain Gugger's avatar
Sylvain Gugger committed
184
        output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202

        # 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).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 attn_mask
        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
        attn_mask = tf.concat([attn_mask, tf.ones((shape_list(attn_mask)[0], 1), dtype=tf.int32)], axis=1)

        # get two different outputs
Sylvain Gugger's avatar
Sylvain Gugger committed
203
204
        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"]
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220

        # select random slice
        random_slice_idx = int(ids_tensor((1,), shape_list(output_from_past)[-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-12)

    def create_and_check_gpt2_lm_head(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
        model = TFGPT2LMHeadModel(config=config)
        inputs = {
            "input_ids": input_ids,
            "attention_mask": input_mask,
            "token_type_ids": token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
221
        result = model(inputs)
222
        self.parent.assertListEqual(
Sylvain Gugger's avatar
Sylvain Gugger committed
223
            list(result["logits"].shape), [self.batch_size, self.seq_length, self.vocab_size],
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
        )

    def create_and_check_gpt2_double_head(
        self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args
    ):
        model = TFGPT2DoubleHeadsModel(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,
            "mc_token_ids": mc_token_ids,
            "attention_mask": multiple_choice_input_mask,
            "token_type_ids": multiple_choice_token_type_ids,
        }
Sylvain Gugger's avatar
Sylvain Gugger committed
241
        result = model(inputs)
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
        self.parent.assertListEqual(
            list(result["lm_logits"].shape), [self.batch_size, self.num_choices, self.seq_length, self.vocab_size],
        )
        self.parent.assertListEqual(list(result["mc_logits"].shape), [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,
            "attention_mask": input_mask,
        }
        return config, inputs_dict


270
@require_tf
271
class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
thomwolf's avatar
thomwolf committed
272

273
    all_model_classes = (TFGPT2Model, TFGPT2LMHeadModel, TFGPT2DoubleHeadsModel) if is_tf_available() else ()
274
    all_generative_model_classes = (TFGPT2LMHeadModel,) if is_tf_available() else ()
thomwolf's avatar
thomwolf committed
275
276

    def setUp(self):
277
        self.model_tester = TFGPT2ModelTester(self)
278
        self.config_tester = ConfigTester(self, config_class=GPT2Config, n_embd=37)
thomwolf's avatar
thomwolf committed
279
280
281
282
283
284
285
286

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

    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)

287
288
289
290
291
292
293
294
    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)

thomwolf's avatar
thomwolf committed
295
296
297
298
299
300
301
302
    def test_gpt2_lm_head(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_gpt2_lm_head(*config_and_inputs)

    def test_gpt2_double_head(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_gpt2_double_head(*config_and_inputs)

303
    @slow
thomwolf's avatar
thomwolf committed
304
    def test_model_from_pretrained(self):
305
        for model_name in TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
306
            model = TFGPT2Model.from_pretrained(model_name)
thomwolf's avatar
thomwolf committed
307
            self.assertIsNotNone(model)
308
309


310
@require_tf
311
class TFGPT2ModelLanguageGenerationTest(unittest.TestCase):
patrickvonplaten's avatar
patrickvonplaten committed
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
    @slow
    def test_lm_generate_gpt2(self):
        model = TFGPT2LMHeadModel.from_pretrained("gpt2")
        input_ids = tf.convert_to_tensor([[464, 3290]], dtype=tf.int32)  # The dog
        expected_output_ids = [
            464,
            3290,
            373,
            1043,
            287,
            257,
            2214,
            1474,
            262,
            16246,
            286,
            2688,
            290,
            2688,
            27262,
            13,
            198,
            198,
            464,
            3290,
        ]  # 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)
339
        self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)
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

    @slow
    def test_lm_generate_distilgpt2(self):
        model = TFGPT2LMHeadModel.from_pretrained("distilgpt2")
        input_ids = tf.convert_to_tensor([[464, 1893]], dtype=tf.int32)  # The president
        expected_output_ids = [
            464,
            1893,
            286,
            262,
            1578,
            1829,
            11,
            290,
            262,
            1893,
            286,
            262,
            1578,
            7526,
            11,
            423,
            587,
            287,
            262,
            2635,
        ]  # The president of the United States, and the president of the United Kingdom, have been in the White

patrickvonplaten's avatar
patrickvonplaten committed
368
        output_ids = model.generate(input_ids, do_sample=False)
369
        self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)