bert_squad_benchmark.py 11.6 KB
Newer Older
davidmochen's avatar
davidmochen committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# 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.
# ==============================================================================
"""Executes BERT SQuAD benchmarks and accuracy tests."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import json
import os
import time

# pylint: disable=g-bad-import-order
from absl import flags
from absl.testing import flagsaver
import tensorflow as tf
# pylint: enable=g-bad-import-order

from official.bert import run_squad
from official.bert.benchmark import benchmark_utils
33
from official.bert.benchmark import squad_evaluate_v1_1
davidmochen's avatar
davidmochen committed
34
35
36
from official.utils.misc import distribution_utils

# pylint: disable=line-too-long
37
PRETRAINED_CHECKPOINT_PATH = 'gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16/bert_model.ckpt'
davidmochen's avatar
davidmochen committed
38
39
40
41
SQUAD_TRAIN_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_train.tf_record'
SQUAD_PREDICT_FILE = 'gs://tf-perfzero-data/bert/squad/dev-v1.1.json'
SQUAD_VOCAB_FILE = 'gs://tf-perfzero-data/bert/squad/vocab.txt'
SQUAD_SMALL_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_small_meta_data'
42
SQUAD_FULL_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/squad/squad_full_meta_data'
davidmochen's avatar
davidmochen committed
43
44
45
46
47
48
49
50
51
MODEL_CONFIG_FILE_PATH = 'gs://cloud-tpu-checkpoints/bert/tf_20/uncased_L-24_H-1024_A-16/bert_config'
# pylint: enable=line-too-long

FLAGS = flags.FLAGS


class BertSquadBenchmarkBase(benchmark_utils.BertBenchmarkBase):
  """Base class to hold methods common to test classes in the module."""

52
53
54
55
56
  def _read_training_summary_from_file(self):
    """Reads the training summary from a file."""
    summary_path = os.path.join(FLAGS.model_dir, 'training_summary.txt')
    with tf.io.gfile.GFile(summary_path, 'rb') as reader:
      return json.loads(reader.read().decode('utf-8'))
57

58
59
60
61
  def _read_input_meta_data_from_file(self):
    """Reads the input metadata from a file."""
    with tf.io.gfile.GFile(FLAGS.input_meta_data_path, 'rb') as reader:
      return json.loads(reader.read().decode('utf-8'))
62

63
64
  def _read_predictions_dataset_from_file(self):
    """Reads the predictions dataset from a file."""
65
66
    with tf.io.gfile.GFile(SQUAD_PREDICT_FILE, 'r') as reader:
      dataset_json = json.load(reader)
67
      return dataset_json['data']
68

69
70
71
  def _read_predictions_from_file(self):
    """Reads the predictions from a file."""
    predictions_file = os.path.join(FLAGS.model_dir, 'predictions.json')
72
    with tf.io.gfile.GFile(predictions_file, 'r') as reader:
73
      return json.load(reader)
74

75
  def _get_distribution_strategy(self, use_ds=True):
76
77
    """Gets the distribution strategy."""
    return distribution_utils.get_distribution_strategy(
78
79
        distribution_strategy='mirrored' if use_ds else 'off',
        num_gpus=self.num_gpus)
80

davidmochen's avatar
davidmochen committed
81
  @flagsaver.flagsaver
82
  def _train_squad(self, use_ds=True, run_eagerly=False):
83
84
    """Runs BERT SQuAD training."""
    input_meta_data = self._read_input_meta_data_from_file()
85
    strategy = self._get_distribution_strategy(use_ds)
davidmochen's avatar
davidmochen committed
86
87
88
89

    run_squad.train_squad(
        strategy=strategy,
        input_meta_data=input_meta_data,
90
        run_eagerly=run_eagerly,
davidmochen's avatar
davidmochen committed
91
        custom_callbacks=[self.timer_callback])
92
93

  @flagsaver.flagsaver
94
  def _evaluate_squad(self, use_ds=True):
95
96
    """Runs BERT SQuAD evaluation."""
    input_meta_data = self._read_input_meta_data_from_file()
97
    strategy = self._get_distribution_strategy(use_ds)
98

99
    run_squad.predict_squad(strategy=strategy, input_meta_data=input_meta_data)
100
101
102
103
104

    dataset = self._read_predictions_dataset_from_file()
    predictions = self._read_predictions_from_file()

    eval_metrics = squad_evaluate_v1_1.evaluate(dataset, predictions)
105
106
    # Use F1 score as reported evaluation metric.
    self.eval_metrics = eval_metrics['f1']
davidmochen's avatar
davidmochen committed
107
108


109
class BertSquadBenchmarkReal(BertSquadBenchmarkBase):
davidmochen's avatar
davidmochen committed
110
111
112
113
114
115
116
117
  """Short benchmark performance tests for BERT SQuAD model.

  Tests BERT SQuAD performance in different GPU configurations.
  The naming convention of below test cases follow
  `benchmark_(number of gpus)_gpu` format.
  """

  def __init__(self, output_dir=None, **kwargs):
118
    super(BertSquadBenchmarkReal, self).__init__(output_dir=output_dir)
davidmochen's avatar
davidmochen committed
119
120

  def _setup(self):
121
122
    """Sets up the benchmark and SQuAD flags."""
    super(BertSquadBenchmarkReal, self)._setup()
davidmochen's avatar
davidmochen committed
123
124
125
126
127
128
    FLAGS.train_data_path = SQUAD_TRAIN_DATA_PATH
    FLAGS.predict_file = SQUAD_PREDICT_FILE
    FLAGS.vocab_file = SQUAD_VOCAB_FILE
    FLAGS.input_meta_data_path = SQUAD_SMALL_INPUT_META_DATA_PATH
    FLAGS.bert_config_file = MODEL_CONFIG_FILE_PATH
    FLAGS.num_train_epochs = 1
129
    FLAGS.steps_per_loop = 1
davidmochen's avatar
davidmochen committed
130

131
132
133
  def _run_and_report_benchmark(self,
                                use_ds=True,
                                run_eagerly=False):
134
    """Runs the benchmark and reports various metrics."""
135
    start_time_sec = time.time()
136
    self._train_squad(use_ds=use_ds, run_eagerly=run_eagerly)
137
138
139
140
141
142
143
144
145
    wall_time_sec = time.time() - start_time_sec

    summary = self._read_training_summary_from_file()

    super(BertSquadBenchmarkReal, self)._report_benchmark(
        stats=summary,
        wall_time_sec=wall_time_sec,
        min_accuracy=0,
        max_accuracy=1)
davidmochen's avatar
davidmochen committed
146
147

  def benchmark_1_gpu(self):
148
    """Tests BERT SQuAD model performance with 1 GPU."""
davidmochen's avatar
davidmochen committed
149
150
151
152
153
154

    self._setup()
    self.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad')
    FLAGS.train_batch_size = 4

155
    self._run_and_report_benchmark()
davidmochen's avatar
davidmochen committed
156

157
158
159
160
161
162
163
  def benchmark_1_gpu_xla(self):
    """Tests BERT SQuAD model performance with 1 GPU with XLA."""

    self._setup()
    self.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_xla_squad')
    FLAGS.train_batch_size = 4
164
    FLAGS.enable_xla = True
165

166
    self._run_and_report_benchmark()
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188

  def benchmark_1_gpu_no_dist_strat(self):
    """Tests BERT SQuAD model performance with 1 GPU without DS."""

    self._setup()
    self.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat_squad')
    FLAGS.train_batch_size = 4

    self._run_and_report_benchmark(use_ds=False)

  def benchmark_1_gpu_eager_no_dist_strat(self):
    """Tests BERT SQuAD model performance with 1 GPU with eager execution."""

    self._setup()
    self.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_1_gpu_eager_no_dist_strat_squad')
    FLAGS.train_batch_size = 4

    self._run_and_report_benchmark(use_ds=False, run_eagerly=True)

davidmochen's avatar
davidmochen committed
189
  def benchmark_2_gpu(self):
190
    """Tests BERT SQuAD model performance with 2 GPUs."""
davidmochen's avatar
davidmochen committed
191
192
193
194
195
196

    self._setup()
    self.num_gpus = 2
    FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu_squad')
    FLAGS.train_batch_size = 8

197
    self._run_and_report_benchmark()
davidmochen's avatar
davidmochen committed
198
199

  def benchmark_4_gpu(self):
200
    """Tests BERT SQuAD model performance with 4 GPUs."""
davidmochen's avatar
davidmochen committed
201
202
203
204
205
206

    self._setup()
    self.num_gpus = 4
    FLAGS.model_dir = self._get_model_dir('benchmark_4_gpu_squad')
    FLAGS.train_batch_size = 16

207
    self._run_and_report_benchmark()
davidmochen's avatar
davidmochen committed
208
209

  def benchmark_8_gpu(self):
210
211
212
213
214
215
216
    """Tests BERT SQuAD model performance with 8 GPUs."""

    self._setup()
    self.num_gpus = 8
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad')
    FLAGS.train_batch_size = 32

217
    self._run_and_report_benchmark()
218

219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
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
  def benchmark_1_gpu_fp16(self):
    """Tests BERT SQuAD model performance with 1 GPU and FP16."""

    self._setup()
    self.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad_fp16')
    FLAGS.train_batch_size = 4
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 'dynamic'

    self._run_and_report_benchmark()

  def benchmark_2_gpu_fp16(self):
    """Tests BERT SQuAD model performance with 2 GPUs and FP16."""

    self._setup()
    self.num_gpus = 2
    FLAGS.model_dir = self._get_model_dir('benchmark_2_gpu_squad_fp16')
    FLAGS.train_batch_size = 8
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 'dynamic'

    self._run_and_report_benchmark()

  def benchmark_4_gpu_fp16(self):
    """Tests BERT SQuAD model performance with 4 GPUs and FP16."""

    self._setup()
    self.num_gpus = 4
    FLAGS.model_dir = self._get_model_dir('benchmark_4_gpu_squad_fp16')
    FLAGS.train_batch_size = 16
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 'dynamic'

    self._run_and_report_benchmark()

  def benchmark_8_gpu_fp16(self):
    """Tests BERT SQuAD model performance with 8 GPUs."""

    self._setup()
    self.num_gpus = 8
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_fp16')
    FLAGS.train_batch_size = 32
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 'dynamic'

    self._run_and_report_benchmark()

267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287

class BertSquadAccuracy(BertSquadBenchmarkBase):
  """Short accuracy test for BERT SQuAD model.

  Tests BERT SQuAD accuracy. The naming convention of below test cases follow
  `benchmark_(number of gpus)_gpu` format.
  """

  def __init__(self, output_dir=None, **kwargs):
    super(BertSquadAccuracy, self).__init__(output_dir=output_dir)

  def _setup(self):
    """Sets up the benchmark and SQuAD flags."""
    super(BertSquadAccuracy, self)._setup()
    FLAGS.train_data_path = SQUAD_TRAIN_DATA_PATH
    FLAGS.predict_file = SQUAD_PREDICT_FILE
    FLAGS.vocab_file = SQUAD_VOCAB_FILE
    FLAGS.input_meta_data_path = SQUAD_FULL_INPUT_META_DATA_PATH
    FLAGS.bert_config_file = MODEL_CONFIG_FILE_PATH
    FLAGS.init_checkpoint = PRETRAINED_CHECKPOINT_PATH
    FLAGS.num_train_epochs = 2
288
    FLAGS.steps_per_loop = 1
289

290
291
292
  def _run_and_report_benchmark(self,
                                use_ds=True,
                                run_eagerly=False):
293
    """Runs the benchmark and reports various metrics."""
294
    start_time_sec = time.time()
295
    self._train_squad(use_ds=use_ds, run_eagerly=run_eagerly)
296
297
298
299
300
301
302
303
304
    self._evaluate_squad()
    wall_time_sec = time.time() - start_time_sec

    summary = self._read_training_summary_from_file()
    summary['eval_metrics'] = self.eval_metrics

    super(BertSquadAccuracy, self)._report_benchmark(
        stats=summary,
        wall_time_sec=wall_time_sec,
305
306
        min_accuracy=0.900,
        max_accuracy=0.908)
307

308
309
310
311
312
313
314
315
316
317
  def benchmark_1_gpu_eager(self):
    """Tests BERT SQuAD model accuracy with 1 GPU with eager execution."""

    self._setup()
    self.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_squad_eager')
    FLAGS.train_batch_size = 4

    self._run_and_report_benchmark(use_ds=False, run_eagerly=True)

318
319
  def benchmark_8_gpu(self):
    """Tests BERT SQuAD model accuracy with 8 GPUs."""
davidmochen's avatar
davidmochen committed
320
321
322
323
324
325

    self._setup()
    self.num_gpus = 8
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad')
    FLAGS.train_batch_size = 32

326
    self._run_and_report_benchmark()
davidmochen's avatar
davidmochen committed
327

328
329
330
331
332
333
334
335
336
337
338
339
  def benchmark_8_gpu_fp16(self):
    """Tests BERT SQuAD model accuracy with 8 GPUs and FP16."""

    self._setup()
    self.num_gpus = 8
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_fp16')
    FLAGS.train_batch_size = 32
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 'dynamic'

    self._run_and_report_benchmark()

340
341
342
343
344
345
346
  def benchmark_8_gpu_xla(self):
    """Tests BERT SQuAD model accuracy with 8 GPUs."""

    self._setup()
    self.num_gpus = 8
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_squad_xla')
    FLAGS.train_batch_size = 32
347
    FLAGS.enable_xla = True
348

349
    self._run_and_report_benchmark()
350

davidmochen's avatar
davidmochen committed
351
352
353

if __name__ == '__main__':
  tf.test.main()