test_pytorch_examples.py 20.8 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 json
18
import logging
19
import os
Aymeric Augustin's avatar
Aymeric Augustin committed
20
import sys
Aymeric Augustin's avatar
Aymeric Augustin committed
21
from unittest.mock import patch
Aymeric Augustin's avatar
Aymeric Augustin committed
22

Stas Bekman's avatar
Stas Bekman committed
23
24
import torch

25
from transformers import ViTMAEForPreTraining, Wav2Vec2ForPreTraining
26
from transformers.testing_utils import CaptureLogger, TestCasePlus, get_gpu_count, slow, torch_device
27
from transformers.utils import is_apex_available
28

29
30
31

SRC_DIRS = [
    os.path.join(os.path.dirname(__file__), dirname)
32
33
34
35
36
    for dirname in [
        "text-generation",
        "text-classification",
        "token-classification",
        "language-modeling",
37
        "multiple-choice",
38
        "question-answering",
Sylvain Gugger's avatar
Sylvain Gugger committed
39
40
        "summarization",
        "translation",
41
        "image-classification",
42
        "speech-recognition",
43
        "audio-classification",
44
        "speech-pretraining",
45
        "image-pretraining",
46
        "semantic-segmentation",
47
    ]
48
49
50
51
52
]
sys.path.extend(SRC_DIRS)


if SRC_DIRS is not None:
53
    import run_audio_classification
Sylvain Gugger's avatar
Sylvain Gugger committed
54
    import run_clm
55
56
    import run_generation
    import run_glue
57
    import run_image_classification
58
    import run_mae
59
    import run_mlm
60
    import run_ner
Sylvain Gugger's avatar
Sylvain Gugger committed
61
    import run_qa as run_squad
62
    import run_semantic_segmentation
63
    import run_seq2seq_qa as run_squad_seq2seq
64
    import run_speech_recognition_ctc
65
    import run_speech_recognition_ctc_adapter
66
    import run_speech_recognition_seq2seq
67
    import run_summarization
68
    import run_swag
69
    import run_translation
70
    import run_wav2vec2_pretraining_no_trainer
Aymeric Augustin's avatar
Aymeric Augustin committed
71

72

73
74
75
logging.basicConfig(level=logging.DEBUG)

logger = logging.getLogger()
76

77

78
79
80
81
82
83
84
85
86
87
88
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


89
def is_cuda_and_apex_available():
90
91
92
93
    is_using_cuda = torch.cuda.is_available() and torch_device == "cuda"
    return is_using_cuda and is_apex_available()


94
95
96
97
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)


98
class ExamplesTests(TestCasePlus):
99
    def test_run_glue(self):
100
101
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
102
            run_glue.py
103
            --model_name_or_path distilbert-base-uncased
104
105
            --output_dir {tmp_dir}
            --overwrite_output_dir
Sylvain Gugger's avatar
Sylvain Gugger committed
106
107
            --train_file ./tests/fixtures/tests_samples/MRPC/train.csv
            --validation_file ./tests/fixtures/tests_samples/MRPC/dev.csv
108
109
            --do_train
            --do_eval
110
111
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
112
113
114
115
116
            --learning_rate=1e-4
            --max_steps=10
            --warmup_steps=2
            --seed=42
            --max_seq_length=128
117
            """.split()
118

119
        if is_cuda_and_apex_available():
120
            testargs.append("--fp16")
121

122
        with patch.object(sys, "argv", testargs):
123
124
            run_glue.main()
            result = get_results(tmp_dir)
125
            self.assertGreaterEqual(result["eval_accuracy"], 0.75)
126

Sylvain Gugger's avatar
Sylvain Gugger committed
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
    def test_run_clm(self):
        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

148
149
        if torch_device == "cpu":
            testargs.append("--use_cpu")
Sylvain Gugger's avatar
Sylvain Gugger committed
150
151

        with patch.object(sys, "argv", testargs):
152
153
            run_clm.main()
            result = get_results(tmp_dir)
Sylvain Gugger's avatar
Sylvain Gugger committed
154
155
            self.assertLess(result["perplexity"], 100)

156
157
158
159
160
161
162
163
164
165
166
167
168
169
    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()

170
171
        if torch_device == "cpu":
            testargs.append("--use_cpu")
172
173
174
175
176
177
178
179
180

        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)

181
    def test_run_mlm(self):
182
183
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
184
            run_mlm.py
Julien Chaumond's avatar
Julien Chaumond committed
185
            --model_name_or_path distilroberta-base
186
187
            --train_file ./tests/fixtures/sample_text.txt
            --validation_file ./tests/fixtures/sample_text.txt
188
            --output_dir {tmp_dir}
Julien Chaumond's avatar
Julien Chaumond committed
189
190
191
            --overwrite_output_dir
            --do_train
            --do_eval
192
            --prediction_loss_only
Julien Chaumond's avatar
Julien Chaumond committed
193
            --num_train_epochs=1
194
        """.split()
195

196
197
        if torch_device == "cpu":
            testargs.append("--use_cpu")
198

Julien Chaumond's avatar
Julien Chaumond committed
199
        with patch.object(sys, "argv", testargs):
200
201
            run_mlm.main()
            result = get_results(tmp_dir)
202
            self.assertLess(result["perplexity"], 42)
Julien Chaumond's avatar
Julien Chaumond committed
203

204
    def test_run_ner(self):
205
206
207
        # 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

208
209
210
211
212
213
214
215
216
217
218
219
        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
220
221
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=2
222
            --num_train_epochs={epochs}
223
            --seed 7
224
225
        """.split()

226
227
        if torch_device == "cpu":
            testargs.append("--use_cpu")
228
229

        with patch.object(sys, "argv", testargs):
230
231
            run_ner.main()
            result = get_results(tmp_dir)
232
            self.assertGreaterEqual(result["eval_accuracy"], 0.75)
233
234
            self.assertLess(result["eval_loss"], 0.5)

235
    def test_run_squad(self):
236
237
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
Russell Klopfer's avatar
Russell Klopfer committed
238
            run_qa.py
Sylvain Gugger's avatar
Sylvain Gugger committed
239
240
241
242
            --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
243
244
            --output_dir {tmp_dir}
            --overwrite_output_dir
245
246
247
248
249
            --max_steps=10
            --warmup_steps=2
            --do_train
            --do_eval
            --learning_rate=2e-4
Sylvain Gugger's avatar
Sylvain Gugger committed
250
251
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
252
253
        """.split()

254
        with patch.object(sys, "argv", testargs):
255
256
            run_squad.main()
            result = get_results(tmp_dir)
Russell Klopfer's avatar
Russell Klopfer committed
257
258
            self.assertGreaterEqual(result["eval_f1"], 30)
            self.assertGreaterEqual(result["eval_exact"], 30)
259

260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
    def test_run_squad_seq2seq(self):
        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)
286
287
            self.assertGreaterEqual(result["eval_f1"], 30)
            self.assertGreaterEqual(result["eval_exact"], 30)
288

289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
    def test_run_swag(self):
        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):
308
309
            run_swag.main()
            result = get_results(tmp_dir)
310
311
            self.assertGreaterEqual(result["eval_accuracy"], 0.8)

312
    def test_generation(self):
313
        testargs = ["run_generation.py", "--prompt=Hello", "--length=10", "--seed=42"]
314

315
        if is_cuda_and_apex_available():
316
317
318
319
320
321
            testargs.append("--fp16")

        model_type, model_name = (
            "--model_type=gpt2",
            "--model_name_or_path=sshleifer/tiny-gpt2",
        )
322
        with patch.object(sys, "argv", testargs + [model_type, model_name]):
323
            result = run_generation.main()
324
            self.assertGreaterEqual(len(result[0]), 10)
325
326

    @slow
327
    def test_run_summarization(self):
328
329
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
330
            run_summarization.py
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
            --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):
347
            run_summarization.main()
348
            result = get_results(tmp_dir)
349
350
351
352
353
354
            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
355
    def test_run_translation(self):
356
357
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
358
            run_translation.py
359
            --model_name_or_path sshleifer/student_marian_en_ro_6_1
360
361
            --source_lang en
            --target_lang ro
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
            --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):
379
            run_translation.main()
380
            result = get_results(tmp_dir)
381
            self.assertGreaterEqual(result["eval_bleu"], 30)
382
383
384
385
386
387
388

    def test_run_image_classification(self):
        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
389
            --dataset_name hf-internal-testing/cats_vs_dogs_sample
390
391
            --do_train
            --do_eval
392
            --learning_rate 1e-4
393
394
395
396
397
398
            --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
399
            --max_steps 10
400
            --train_val_split 0.1
401
            --seed 42
402
403
404
405
406
407
408
409
410
        """.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)
411
412
413
414
415
416
417

    def test_run_speech_recognition_ctc(self):
        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
418
            --dataset_name hf-internal-testing/librispeech_asr_dummy
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
            --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"])
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472

    def test_run_speech_recognition_ctc_adapter(self):
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_speech_recognition_ctc_adapter.py
            --output_dir {tmp_dir}
            --model_name_or_path hf-internal-testing/tiny-random-wav2vec2
            --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 1
            --remove_unused_columns False
            --overwrite_output_dir True
            --preprocessing_num_workers 16
            --max_steps 10
            --target_language tur
            --seed 42
        """.split()

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

        with patch.object(sys, "argv", testargs):
            run_speech_recognition_ctc_adapter.main()
            result = get_results(tmp_dir)
            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, "./adapter.tur.safetensors")))
            self.assertLess(result["eval_loss"], result["train_loss"])
473

474
475
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
    def test_run_speech_recognition_seq2seq(self):
        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"])

504
505
506
507
508
509
510
511
512
513
    def test_run_audio_classification(self):
        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
514
            --audio_column_name audio
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
            --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"])
535
536
537
538
539
540
541

    def test_run_wav2vec2_pretraining(self):
        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
542
            --dataset_name hf-internal-testing/librispeech_asr_dummy
543
544
545
            --dataset_config_names clean
            --dataset_split_names validation
            --learning_rate 1e-4
546
547
            --per_device_train_batch_size 4
            --per_device_eval_batch_size 4
548
            --preprocessing_num_workers 16
549
            --max_train_steps 2
550
551
552
553
554
555
556
557
558
559
560
            --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)
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588

    def test_run_vit_mae_pretraining(self):
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_mae.py
            --output_dir {tmp_dir}
            --dataset_name hf-internal-testing/cats_vs_dogs_sample
            --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
            --dataloader_num_workers 16
            --metric_for_best_model accuracy
            --max_steps 10
            --train_val_split 0.1
            --seed 42
        """.split()

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

        with patch.object(sys, "argv", testargs):
            run_mae.main()
            model = ViTMAEForPreTraining.from_pretrained(tmp_dir)
            self.assertIsNotNone(model)
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613

    def test_run_semantic_segmentation(self):
        tmp_dir = self.get_auto_remove_tmp_dir()
        testargs = f"""
            run_semantic_segmentation.py
            --output_dir {tmp_dir}
            --dataset_name huggingface/semantic-segmentation-test-sample
            --do_train
            --do_eval
            --remove_unused_columns False
            --overwrite_output_dir True
            --max_steps 10
            --learning_rate=2e-4
            --per_device_train_batch_size=2
            --per_device_eval_batch_size=1
            --seed 32
        """.split()

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

        with patch.object(sys, "argv", testargs):
            run_semantic_segmentation.main()
            result = get_results(tmp_dir)
            self.assertGreaterEqual(result["eval_overall_accuracy"], 0.1)