test_examples.py 19.2 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2018 HuggingFace Inc..
#
# 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 argparse
18
import json
19
import logging
20
import os
Aymeric Augustin's avatar
Aymeric Augustin committed
21
import sys
Aymeric Augustin's avatar
Aymeric Augustin committed
22
from unittest.mock import patch
Aymeric Augustin's avatar
Aymeric Augustin committed
23

Stas Bekman's avatar
Stas Bekman committed
24
25
import torch

26
from transformers import Wav2Vec2ForPreTraining
27
from transformers.file_utils import is_apex_available
28
from transformers.testing_utils import CaptureLogger, TestCasePlus, get_gpu_count, slow, torch_device
29

30
31
32

SRC_DIRS = [
    os.path.join(os.path.dirname(__file__), dirname)
33
34
35
36
37
    for dirname in [
        "text-generation",
        "text-classification",
        "token-classification",
        "language-modeling",
38
        "multiple-choice",
39
        "question-answering",
Sylvain Gugger's avatar
Sylvain Gugger committed
40
41
        "summarization",
        "translation",
42
        "image-classification",
43
        "speech-recognition",
44
        "audio-classification",
45
        "speech-pretraining",
46
    ]
47
48
49
50
51
]
sys.path.extend(SRC_DIRS)


if SRC_DIRS is not None:
52
    import run_audio_classification
Sylvain Gugger's avatar
Sylvain Gugger committed
53
    import run_clm
54
55
    import run_generation
    import run_glue
56
    import run_image_classification
57
    import run_mlm
58
    import run_ner
Sylvain Gugger's avatar
Sylvain Gugger committed
59
    import run_qa as run_squad
60
    import run_seq2seq_qa as run_squad_seq2seq
61
    import run_speech_recognition_ctc
62
    import run_speech_recognition_seq2seq
63
    import run_summarization
64
    import run_swag
65
    import run_translation
66
    import run_wav2vec2_pretraining_no_trainer
Aymeric Augustin's avatar
Aymeric Augustin committed
67

68

69
70
71
logging.basicConfig(level=logging.DEBUG)

logger = logging.getLogger()
72

73

74
75
def get_setup_file():
    parser = argparse.ArgumentParser()
76
    parser.add_argument("-f")
77
78
79
80
    args = parser.parse_args()
    return args.f


81
82
83
84
85
86
87
88
89
90
91
def get_results(output_dir):
    results = {}
    path = os.path.join(output_dir, "all_results.json")
    if os.path.exists(path):
        with open(path, "r") as f:
            results = json.load(f)
    else:
        raise ValueError(f"can't find {path}")
    return results


92
def is_cuda_and_apex_available():
93
94
95
96
    is_using_cuda = torch.cuda.is_available() and torch_device == "cuda"
    return is_using_cuda and is_apex_available()


97
class ExamplesTests(TestCasePlus):
98
99
100
101
    def test_run_glue(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

102
103
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
104
            run_glue.py
105
            --model_name_or_path distilbert-base-uncased
106
107
            --output_dir {tmp_dir}
            --overwrite_output_dir
Sylvain Gugger's avatar
Sylvain Gugger committed
108
109
            --train_file ./tests/fixtures/tests_samples/MRPC/train.csv
            --validation_file ./tests/fixtures/tests_samples/MRPC/dev.csv
110
111
            --do_train
            --do_eval
112
113
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
114
115
116
117
118
            --learning_rate=1e-4
            --max_steps=10
            --warmup_steps=2
            --seed=42
            --max_seq_length=128
119
            """.split()
120

121
        if is_cuda_and_apex_available():
122
            testargs.append("--fp16")
123

124
        with patch.object(sys, "argv", testargs):
125
126
            run_glue.main()
            result = get_results(tmp_dir)
127
            self.assertGreaterEqual(result["eval_accuracy"], 0.75)
128

Sylvain Gugger's avatar
Sylvain Gugger committed
129
130
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
    def test_run_clm(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_clm.py
            --model_name_or_path distilgpt2
            --train_file ./tests/fixtures/sample_text.txt
            --validation_file ./tests/fixtures/sample_text.txt
            --do_train
            --do_eval
            --block_size 128
            --per_device_train_batch_size 5
            --per_device_eval_batch_size 5
            --num_train_epochs 2
            --output_dir {tmp_dir}
            --overwrite_output_dir
            """.split()

        if torch.cuda.device_count() > 1:
            # Skipping because there are not enough batches to train the model + would need a drop_last to work.
            return

        if torch_device != "cuda":
            testargs.append("--no_cuda")

        with patch.object(sys, "argv", testargs):
157
158
            run_clm.main()
            result = get_results(tmp_dir)
Sylvain Gugger's avatar
Sylvain Gugger committed
159
160
            self.assertLess(result["perplexity"], 100)

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
    def test_run_clm_config_overrides(self):
        # test that config_overrides works, despite the misleading dumps of default un-updated
        # config via tokenizer

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_clm.py
            --model_type gpt2
            --tokenizer_name gpt2
            --train_file ./tests/fixtures/sample_text.txt
            --output_dir {tmp_dir}
            --config_overrides n_embd=10,n_head=2
            """.split()

        if torch_device != "cuda":
            testargs.append("--no_cuda")

        logger = run_clm.logger
        with patch.object(sys, "argv", testargs):
            with CaptureLogger(logger) as cl:
                run_clm.main()

        self.assertIn('"n_embd": 10', cl.out)
        self.assertIn('"n_head": 2', cl.out)

186
    def test_run_mlm(self):
Julien Chaumond's avatar
Julien Chaumond committed
187
188
189
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

190
191
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
192
            run_mlm.py
Julien Chaumond's avatar
Julien Chaumond committed
193
            --model_name_or_path distilroberta-base
194
195
            --train_file ./tests/fixtures/sample_text.txt
            --validation_file ./tests/fixtures/sample_text.txt
196
            --output_dir {tmp_dir}
Julien Chaumond's avatar
Julien Chaumond committed
197
198
199
            --overwrite_output_dir
            --do_train
            --do_eval
200
            --prediction_loss_only
Julien Chaumond's avatar
Julien Chaumond committed
201
            --num_train_epochs=1
202
        """.split()
203
204
205

        if torch_device != "cuda":
            testargs.append("--no_cuda")
206

Julien Chaumond's avatar
Julien Chaumond committed
207
        with patch.object(sys, "argv", testargs):
208
209
            run_mlm.main()
            result = get_results(tmp_dir)
210
            self.assertLess(result["perplexity"], 42)
Julien Chaumond's avatar
Julien Chaumond committed
211

212
213
214
215
    def test_run_ner(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

216
217
218
        # with so little data distributed training needs more epochs to get the score on par with 0/1 gpu
        epochs = 7 if get_gpu_count() > 1 else 2

219
220
221
222
223
224
225
226
227
228
229
230
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_ner.py
            --model_name_or_path bert-base-uncased
            --train_file tests/fixtures/tests_samples/conll/sample.json
            --validation_file tests/fixtures/tests_samples/conll/sample.json
            --output_dir {tmp_dir}
            --overwrite_output_dir
            --do_train
            --do_eval
            --warmup_steps=2
            --learning_rate=2e-4
Sylvain Gugger's avatar
Sylvain Gugger committed
231
232
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=2
233
            --num_train_epochs={epochs}
234
            --seed 7
235
236
237
238
239
240
        """.split()

        if torch_device != "cuda":
            testargs.append("--no_cuda")

        with patch.object(sys, "argv", testargs):
241
242
            run_ner.main()
            result = get_results(tmp_dir)
243
            self.assertGreaterEqual(result["eval_accuracy"], 0.75)
244
245
            self.assertLess(result["eval_loss"], 0.5)

246
247
248
249
    def test_run_squad(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

250
251
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
Russell Klopfer's avatar
Russell Klopfer committed
252
            run_qa.py
Sylvain Gugger's avatar
Sylvain Gugger committed
253
254
255
256
            --model_name_or_path bert-base-uncased
            --version_2_with_negative
            --train_file tests/fixtures/tests_samples/SQUAD/sample.json
            --validation_file tests/fixtures/tests_samples/SQUAD/sample.json
257
258
            --output_dir {tmp_dir}
            --overwrite_output_dir
259
260
261
262
263
            --max_steps=10
            --warmup_steps=2
            --do_train
            --do_eval
            --learning_rate=2e-4
Sylvain Gugger's avatar
Sylvain Gugger committed
264
265
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
266
267
        """.split()

268
        with patch.object(sys, "argv", testargs):
269
270
            run_squad.main()
            result = get_results(tmp_dir)
Russell Klopfer's avatar
Russell Klopfer committed
271
272
            self.assertGreaterEqual(result["eval_f1"], 30)
            self.assertGreaterEqual(result["eval_exact"], 30)
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
    def test_run_squad_seq2seq(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_seq2seq_qa.py
            --model_name_or_path t5-small
            --context_column context
            --question_column question
            --answer_column answers
            --version_2_with_negative
            --train_file tests/fixtures/tests_samples/SQUAD/sample.json
            --validation_file tests/fixtures/tests_samples/SQUAD/sample.json
            --output_dir {tmp_dir}
            --overwrite_output_dir
            --max_steps=10
            --warmup_steps=2
            --do_train
            --do_eval
            --learning_rate=2e-4
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
            --predict_with_generate
        """.split()

        with patch.object(sys, "argv", testargs):
            run_squad_seq2seq.main()
            result = get_results(tmp_dir)
303
304
            self.assertGreaterEqual(result["eval_f1"], 30)
            self.assertGreaterEqual(result["eval_exact"], 30)
305

306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
    def test_run_swag(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_swag.py
            --model_name_or_path bert-base-uncased
            --train_file tests/fixtures/tests_samples/swag/sample.json
            --validation_file tests/fixtures/tests_samples/swag/sample.json
            --output_dir {tmp_dir}
            --overwrite_output_dir
            --max_steps=20
            --warmup_steps=2
            --do_train
            --do_eval
            --learning_rate=2e-4
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
        """.split()

        with patch.object(sys, "argv", testargs):
328
329
            run_swag.main()
            result = get_results(tmp_dir)
330
331
            self.assertGreaterEqual(result["eval_accuracy"], 0.8)

332
333
334
335
    def test_generation(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

336
        testargs = ["run_generation.py", "--prompt=Hello", "--length=10", "--seed=42"]
337

338
        if is_cuda_and_apex_available():
339
340
341
342
343
344
            testargs.append("--fp16")

        model_type, model_name = (
            "--model_type=gpt2",
            "--model_name_or_path=sshleifer/tiny-gpt2",
        )
345
        with patch.object(sys, "argv", testargs + [model_type, model_name]):
346
            result = run_generation.main()
347
            self.assertGreaterEqual(len(result[0]), 10)
348
349

    @slow
350
    def test_run_summarization(self):
351
352
353
354
355
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
356
            run_summarization.py
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
            --model_name_or_path t5-small
            --train_file tests/fixtures/tests_samples/xsum/sample.json
            --validation_file tests/fixtures/tests_samples/xsum/sample.json
            --output_dir {tmp_dir}
            --overwrite_output_dir
            --max_steps=50
            --warmup_steps=8
            --do_train
            --do_eval
            --learning_rate=2e-4
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
            --predict_with_generate
        """.split()

        with patch.object(sys, "argv", testargs):
373
            run_summarization.main()
374
            result = get_results(tmp_dir)
375
376
377
378
379
380
            self.assertGreaterEqual(result["eval_rouge1"], 10)
            self.assertGreaterEqual(result["eval_rouge2"], 2)
            self.assertGreaterEqual(result["eval_rougeL"], 7)
            self.assertGreaterEqual(result["eval_rougeLsum"], 7)

    @slow
381
    def test_run_translation(self):
382
383
384
385
386
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
387
            run_translation.py
388
            --model_name_or_path sshleifer/student_marian_en_ro_6_1
389
390
            --source_lang en
            --target_lang ro
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
            --train_file tests/fixtures/tests_samples/wmt16/sample.json
            --validation_file tests/fixtures/tests_samples/wmt16/sample.json
            --output_dir {tmp_dir}
            --overwrite_output_dir
            --max_steps=50
            --warmup_steps=8
            --do_train
            --do_eval
            --learning_rate=3e-3
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
            --predict_with_generate
            --source_lang en_XX
            --target_lang ro_RO
        """.split()

        with patch.object(sys, "argv", testargs):
408
            run_translation.main()
409
            result = get_results(tmp_dir)
410
            self.assertGreaterEqual(result["eval_bleu"], 30)
411
412
413
414
415
416
417
418
419
420

    def test_run_image_classification(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_image_classification.py
            --output_dir {tmp_dir}
            --model_name_or_path google/vit-base-patch16-224-in21k
421
            --dataset_name hf-internal-testing/cats_vs_dogs_sample
422
423
            --do_train
            --do_eval
424
            --learning_rate 1e-4
425
426
427
428
429
430
            --per_device_train_batch_size 2
            --per_device_eval_batch_size 1
            --remove_unused_columns False
            --overwrite_output_dir True
            --dataloader_num_workers 16
            --metric_for_best_model accuracy
431
            --max_steps 10
432
            --train_val_split 0.1
433
            --seed 42
434
435
436
437
438
439
440
441
442
        """.split()

        if is_cuda_and_apex_available():
            testargs.append("--fp16")

        with patch.object(sys, "argv", testargs):
            run_image_classification.main()
            result = get_results(tmp_dir)
            self.assertGreaterEqual(result["eval_accuracy"], 0.8)
443
444
445
446
447
448
449
450
451
452

    def test_run_speech_recognition_ctc(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_speech_recognition_ctc.py
            --output_dir {tmp_dir}
            --model_name_or_path hf-internal-testing/tiny-random-wav2vec2
Patrick von Platen's avatar
Patrick von Platen committed
453
            --dataset_name hf-internal-testing/librispeech_asr_dummy
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
            --dataset_config_name clean
            --train_split_name validation
            --eval_split_name validation
            --do_train
            --do_eval
            --learning_rate 1e-4
            --per_device_train_batch_size 2
            --per_device_eval_batch_size 1
            --remove_unused_columns False
            --overwrite_output_dir True
            --preprocessing_num_workers 16
            --max_steps 10
            --seed 42
        """.split()

        if is_cuda_and_apex_available():
            testargs.append("--fp16")

        with patch.object(sys, "argv", testargs):
            run_speech_recognition_ctc.main()
            result = get_results(tmp_dir)
            self.assertLess(result["eval_loss"], result["train_loss"])
476

477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
    def test_run_speech_recognition_seq2seq(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_speech_recognition_seq2seq.py
            --output_dir {tmp_dir}
            --model_name_or_path hf-internal-testing/tiny-random-speech-encoder-decoder
            --dataset_name hf-internal-testing/librispeech_asr_dummy
            --dataset_config_name clean
            --train_split_name validation
            --eval_split_name validation
            --do_train
            --do_eval
            --learning_rate 1e-4
            --per_device_train_batch_size 2
            --per_device_eval_batch_size 4
            --remove_unused_columns False
            --overwrite_output_dir True
            --preprocessing_num_workers 16
            --max_steps 10
            --seed 42
        """.split()

        if is_cuda_and_apex_available():
            testargs.append("--fp16")

        with patch.object(sys, "argv", testargs):
            run_speech_recognition_seq2seq.main()
            result = get_results(tmp_dir)
            self.assertLess(result["eval_loss"], result["train_loss"])

510
511
512
513
514
515
516
517
518
519
520
521
522
    def test_run_audio_classification(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_audio_classification.py
            --output_dir {tmp_dir}
            --model_name_or_path hf-internal-testing/tiny-random-wav2vec2
            --dataset_name anton-l/superb_demo
            --dataset_config_name ks
            --train_split_name test
            --eval_split_name test
523
            --audio_column_name audio
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
            --label_column_name label
            --do_train
            --do_eval
            --learning_rate 1e-4
            --per_device_train_batch_size 2
            --per_device_eval_batch_size 1
            --remove_unused_columns False
            --overwrite_output_dir True
            --num_train_epochs 10
            --max_steps 50
            --seed 42
        """.split()

        if is_cuda_and_apex_available():
            testargs.append("--fp16")

        with patch.object(sys, "argv", testargs):
            run_audio_classification.main()
            result = get_results(tmp_dir)
            self.assertLess(result["eval_loss"], result["train_loss"])
544
545
546
547
548
549
550
551
552
553

    def test_run_wav2vec2_pretraining(self):
        stream_handler = logging.StreamHandler(sys.stdout)
        logger.addHandler(stream_handler)

        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_wav2vec2_pretraining_no_trainer.py
            --output_dir {tmp_dir}
            --model_name_or_path hf-internal-testing/tiny-random-wav2vec2
Patrick von Platen's avatar
Patrick von Platen committed
554
            --dataset_name hf-internal-testing/librispeech_asr_dummy
555
556
557
            --dataset_config_names clean
            --dataset_split_names validation
            --learning_rate 1e-4
558
559
            --per_device_train_batch_size 4
            --per_device_eval_batch_size 4
560
            --preprocessing_num_workers 16
561
            --max_train_steps 2
562
563
564
565
566
567
568
569
570
571
572
            --validation_split_percentage 5
            --seed 42
        """.split()

        if is_cuda_and_apex_available():
            testargs.append("--fp16")

        with patch.object(sys, "argv", testargs):
            run_wav2vec2_pretraining_no_trainer.main()
            model = Wav2Vec2ForPreTraining.from_pretrained(tmp_dir)
            self.assertIsNotNone(model)