keras_imagenet_benchmark.py 20.6 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
# Copyright 2018 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 Keras benchmarks and accuracy tests."""
from __future__ import print_function

import os
19
import time
20
21
22
23

from absl import flags

from official.resnet import imagenet_main
Toby Boyd's avatar
Toby Boyd committed
24
from official.resnet.keras import keras_benchmark
25
26
27
from official.resnet.keras import keras_common
from official.resnet.keras import keras_imagenet_main

Toby Boyd's avatar
Toby Boyd committed
28
29
MIN_TOP_1_ACCURACY = 0.76
MAX_TOP_1_ACCURACY = 0.77
30

Toby Boyd's avatar
Toby Boyd committed
31
FLAGS = flags.FLAGS
32
33


Toby Boyd's avatar
Toby Boyd committed
34
35
class Resnet50KerasAccuracy(keras_benchmark.KerasBenchmark):
  """Benchmark accuracy tests for ResNet50 in Keras."""
36

37
  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
38
39
40
41
42
    """A benchmark class.

    Args:
      output_dir: directory where to output e.g. log files
      root_data_dir: directory under which to look for dataset
43
44
45
      **kwargs: arbitrary named arguments. This is needed to make the
                constructor forward compatible in case PerfZero provides more
                named arguments before updating the constructor.
46
47
    """

48
    flag_methods = [
49
50
        keras_common.define_keras_flags,
        lambda: imagenet_main.define_imagenet_flags(dynamic_loss_scale=True)
51
    ]
Toby Boyd's avatar
Toby Boyd committed
52

53
    self.data_dir = os.path.join(root_data_dir, 'imagenet')
54
55
    super(Resnet50KerasAccuracy, self).__init__(
        output_dir=output_dir, flag_methods=flag_methods)
56

Toby Boyd's avatar
Toby Boyd committed
57
  def benchmark_graph_8_gpu(self):
58
59
    """Test Keras model with Keras fit/dist_strat and 8 GPUs."""
    self._setup()
Toby Boyd's avatar
Toby Boyd committed
60
    FLAGS.num_gpus = 8
61
    FLAGS.data_dir = self.data_dir
62
    FLAGS.batch_size = 128 * 8
Toby Boyd's avatar
Toby Boyd committed
63
    FLAGS.train_epochs = 90
64
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
65
    FLAGS.dtype = 'fp32'
66
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
67
68

  def benchmark_8_gpu(self):
69
70
    """Test Keras model with eager, dist_strat and 8 GPUs."""
    self._setup()
Toby Boyd's avatar
Toby Boyd committed
71
    FLAGS.num_gpus = 8
72
    FLAGS.data_dir = self.data_dir
73
    FLAGS.batch_size = 128 * 8
Toby Boyd's avatar
Toby Boyd committed
74
    FLAGS.train_epochs = 90
75
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
76
77
    FLAGS.dtype = 'fp32'
    FLAGS.enable_eager = True
78
79
    # Add some thread tunings to improve performance.
    FLAGS.datasets_num_private_threads = 14
80
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
81

Reed's avatar
Reed committed
82
83
84
85
86
87
88
89
90
91
  def benchmark_8_gpu_fp16(self):
    """Test Keras model with eager, dist_strat, 8 GPUs, and fp16."""
    self._setup()
    FLAGS.num_gpus = 8
    FLAGS.data_dir = self.data_dir
    FLAGS.batch_size = 256 * 8
    FLAGS.train_epochs = 90
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
92
93
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
Reed's avatar
Reed committed
94
95
96
97
98
99
100
101
102
103
104
105
106
    self._run_and_report_benchmark()

  def benchmark_xla_8_gpu_fp16(self):
    """Test Keras model with XLA, eager, dist_strat, 8 GPUs and fp16."""
    self._setup()
    FLAGS.num_gpus = 8
    FLAGS.data_dir = self.data_dir
    FLAGS.batch_size = 256 * 8
    FLAGS.train_epochs = 90
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
107
108
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
Reed's avatar
Reed committed
109
110
    self._run_and_report_benchmark()

111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
  def benchmark_xla_8_gpu_fp16_dynamic(self):
    """Test Keras model with XLA, eager, dist_strat, 8 GPUs, dynamic fp16."""
    self._setup()
    FLAGS.num_gpus = 8
    FLAGS.data_dir = self.data_dir
    FLAGS.batch_size = 256 * 8
    FLAGS.train_epochs = 90
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16_dynamic')
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.loss_scale = 'dynamic'
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

127
128
129
130
131
132
  def _run_and_report_benchmark(self):
    start_time_sec = time.time()
    stats = keras_imagenet_main.run(flags.FLAGS)
    wall_time_sec = time.time() - start_time_sec

    super(Resnet50KerasAccuracy, self)._report_benchmark(
Toby Boyd's avatar
Toby Boyd committed
133
        stats,
134
        wall_time_sec,
Toby Boyd's avatar
Toby Boyd committed
135
136
        top_1_min=MIN_TOP_1_ACCURACY,
        top_1_max=MAX_TOP_1_ACCURACY,
137
        total_batch_size=FLAGS.batch_size,
Toby Boyd's avatar
Toby Boyd committed
138
        log_steps=100)
139
140
141
142

  def _get_model_dir(self, folder_name):
    return os.path.join(self.output_dir, folder_name)

Toby Boyd's avatar
Toby Boyd committed
143
144
145
146
147

class Resnet50KerasBenchmarkBase(keras_benchmark.KerasBenchmark):
  """Resnet50 benchmarks."""

  def __init__(self, output_dir=None, default_flags=None):
148
    flag_methods = [
149
150
        keras_common.define_keras_flags,
        lambda: imagenet_main.define_imagenet_flags(dynamic_loss_scale=True)
151
    ]
Toby Boyd's avatar
Toby Boyd committed
152
153
154
155
156
157

    super(Resnet50KerasBenchmarkBase, self).__init__(
        output_dir=output_dir,
        flag_methods=flag_methods,
        default_flags=default_flags)

158
159
  def _run_and_report_benchmark(self):
    start_time_sec = time.time()
Toby Boyd's avatar
Toby Boyd committed
160
    stats = keras_imagenet_main.run(FLAGS)
161
162
163
164
165
166
167
    wall_time_sec = time.time() - start_time_sec

    super(Resnet50KerasBenchmarkBase, self)._report_benchmark(
        stats,
        wall_time_sec,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)
Toby Boyd's avatar
Toby Boyd committed
168
169

  def benchmark_1_gpu_no_dist_strat(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
170
    """Test Keras model with 1 GPU, no distribution strategy."""
Toby Boyd's avatar
Toby Boyd committed
171
172
173
174
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
175
    FLAGS.distribution_strategy = 'off'
176
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
Toby Boyd's avatar
Toby Boyd committed
177
    FLAGS.batch_size = 128
178
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
179
180

  def benchmark_graph_1_gpu_no_dist_strat(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
181
    """Test Keras model in legacy graph mode with 1 GPU, no dist strat."""
Toby Boyd's avatar
Toby Boyd committed
182
183
184
185
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
186
    FLAGS.distribution_strategy = 'off'
187
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
Toby Boyd's avatar
Toby Boyd committed
188
    FLAGS.batch_size = 128
189
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
190
191

  def benchmark_1_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
192
    """Test Keras model with 1 GPU."""
Toby Boyd's avatar
Toby Boyd committed
193
194
195
196
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
197
    FLAGS.distribution_strategy = 'default'
198
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
Toby Boyd's avatar
Toby Boyd committed
199
    FLAGS.batch_size = 128
200
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
201

Haoyu Zhang's avatar
Haoyu Zhang committed
202
203
204
205
206
207
208
209
210
211
212
213
  def benchmark_xla_1_gpu(self):
    """Test Keras model with XLA and 1 GPU."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu')
    FLAGS.batch_size = 128
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
214
  def benchmark_1_gpu_fp16(self):
215
    """Test Keras model with 1 GPU and fp16."""
Reed's avatar
Reed committed
216
217
218
219
220
221
222
223
224
225
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

226
227
228
229
230
231
232
233
234
235
236
237
238
  def benchmark_1_gpu_fp16_dynamic(self):
    """Test Keras model with 1 GPU, fp16, and dynamic loss scaling."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16_dynamic')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
    FLAGS.loss_scale = 'dynamic'
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
239
240
241
242
243
244
245
246
247
248
249
250
251
  def benchmark_xla_1_gpu_fp16(self):
    """Test Keras model with XLA, 1 GPU and fp16."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

252
253
254
255
256
257
258
259
260
261
262
263
264
265
  def benchmark_xla_1_gpu_fp16_dynamic(self):
    """Test Keras model with XLA, 1 GPU, fp16, and dynamic loss scaling."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_dynamic')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
    FLAGS.loss_scale = 'dynamic'
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
266
  def benchmark_graph_1_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
267
    """Test Keras model in legacy graph mode with 1 GPU."""
Toby Boyd's avatar
Toby Boyd committed
268
269
270
271
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
272
    FLAGS.distribution_strategy = 'default'
273
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
Toby Boyd's avatar
Toby Boyd committed
274
    FLAGS.batch_size = 128
275
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
276

Haoyu Zhang's avatar
Haoyu Zhang committed
277
278
279
280
281
282
283
284
285
286
287
288
  def benchmark_graph_xla_1_gpu(self):
    """Test Keras model in legacy graph mode with XLA and 1 GPU."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu')
    FLAGS.batch_size = 128
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
289
  def benchmark_8_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
290
    """Test Keras model with 8 GPUs."""
Toby Boyd's avatar
Toby Boyd committed
291
292
293
294
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
295
    FLAGS.distribution_strategy = 'default'
296
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
297
    FLAGS.batch_size = 128 * 8  # 8 GPUs
298
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
299

300
  def benchmark_8_gpu_tweaked(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
301
    """Test Keras model with manual config tuning and 8 GPUs."""
302
303
304
305
306
307
308
309
310
311
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_tweaked')
    FLAGS.batch_size = 128 * 8  # 8 GPUs
    FLAGS.datasets_num_private_threads = 14
    self._run_and_report_benchmark()

Haoyu Zhang's avatar
Haoyu Zhang committed
312
313
314
315
316
317
318
319
320
  def benchmark_xla_8_gpu(self):
    """Test Keras model with XLA and 8 GPUs."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu')
321
    FLAGS.batch_size = 128 * 8  # 8 GPUs
Haoyu Zhang's avatar
Haoyu Zhang committed
322
323
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
324
  def benchmark_8_gpu_fp16(self):
325
    """Test Keras model with 8 GPUs and fp16."""
Reed's avatar
Reed committed
326
327
328
    self._setup()

    FLAGS.num_gpus = 8
329
    FLAGS.dtype = 'fp16'
Reed's avatar
Reed committed
330
331
332
333
334
335
    FLAGS.enable_eager = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()

336
337
338
339
340
341
342
343
344
345
346
347
348
  def benchmark_8_gpu_fp16_tweaked(self):
    """Test Keras model with 8 GPUs and fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
  def benchmark_8_gpu_fp16_dynamic_tweaked(self):
    """Test Keras model with 8 GPUs, fp16, and dynamic loss scaling."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_8_gpu_fp16_dynamic_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.loss_scale = 'dynamic'
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
364
  def benchmark_xla_8_gpu_fp16(self):
365
    """Test Keras model with XLA, 8 GPUs and fp16."""
Reed's avatar
Reed committed
366
367
368
    self._setup()

    FLAGS.num_gpus = 8
369
    FLAGS.dtype = 'fp16'
Reed's avatar
Reed committed
370
371
372
373
374
375
376
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()

377
378
379
380
381
382
383
384
385
386
387
388
389
390
  def benchmark_xla_8_gpu_fp16_tweaked(self):
    """Test Keras model with manual config tuning, XLA, 8 GPUs and fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
  def benchmark_xla_8_gpu_fp16_dynamic_tweaked(self):
    """Test Keras model with config tuning, XLA, 8 GPUs and dynamic fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_xla_8_gpu_fp16_dynamic_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.loss_scale = 'dynamic'
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
  def benchmark_xla_8_gpu_fp16_tensorboard_tweaked(self):
    """Test to track Tensorboard performance overhead."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_xla_8_gpu_fp16_tensorboard_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    FLAGS.enable_tensorboard = True
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
423
  def benchmark_graph_8_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
424
    """Test Keras model in legacy graph mode with 8 GPUs."""
Toby Boyd's avatar
Toby Boyd committed
425
426
427
428
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
429
    FLAGS.distribution_strategy = 'default'
430
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
431
    FLAGS.batch_size = 128 * 8  # 8 GPUs
432
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
433

Haoyu Zhang's avatar
Haoyu Zhang committed
434
435
436
437
438
439
440
441
442
  def benchmark_graph_xla_8_gpu(self):
    """Test Keras model in legacy graph mode with XLA and 8 GPUs."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_8_gpu')
443
    FLAGS.batch_size = 128 * 8  # 8 GPUs
Haoyu Zhang's avatar
Haoyu Zhang committed
444
445
    self._run_and_report_benchmark()

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
473
474
475
  def benchmark_graph_xla_8_gpu_fp16(self):
    """Test Keras model in legacy graph mode with XLA, 8 GPUs and fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()

  def benchmark_graph_xla_8_gpu_fp16_tweaked(self):
    """Test Keras model in legacy graph mode with manual config tuning, XLA,
       8 GPUs and fp16.
    """
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_xla_8_gpu_fp16_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
  def benchmark_graph_xla_8_gpu_fp16_dynamic_tweaked(self):
    """Test graph Keras with config tuning, XLA, 8 GPUs and dynamic fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
    FLAGS.distribution_strategy = 'default'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_xla_8_gpu_fp16_dynamic_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.loss_scale = 'dynamic'
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
492
493
494
495
496
497
  def fill_report_object(self, stats):
    super(Resnet50KerasBenchmarkBase, self).fill_report_object(
        stats,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)

Toby Boyd's avatar
Toby Boyd committed
498
499
500
501

class Resnet50KerasBenchmarkSynth(Resnet50KerasBenchmarkBase):
  """Resnet50 synthetic benchmark tests."""

502
  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
Toby Boyd's avatar
Toby Boyd committed
503
504
505
506
507
508
    def_flags = {}
    def_flags['skip_eval'] = True
    def_flags['use_synthetic_data'] = True
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

509
510
    super(Resnet50KerasBenchmarkSynth, self).__init__(
        output_dir=output_dir, default_flags=def_flags)
Toby Boyd's avatar
Toby Boyd committed
511
512
513
514
515


class Resnet50KerasBenchmarkReal(Resnet50KerasBenchmarkBase):
  """Resnet50 real data benchmark tests."""

516
  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
Toby Boyd's avatar
Toby Boyd committed
517
518
    def_flags = {}
    def_flags['skip_eval'] = True
519
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
Toby Boyd's avatar
Toby Boyd committed
520
521
522
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

523
524
    super(Resnet50KerasBenchmarkReal, self).__init__(
        output_dir=output_dir, default_flags=def_flags)
525
526


527
class TrivialKerasBenchmarkReal(keras_benchmark.KerasBenchmark):
528
529
530
531
  """Trivial model with real data benchmark tests."""

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
    flag_methods = [
532
533
        keras_common.define_keras_flags,
        lambda: imagenet_main.define_imagenet_flags(dynamic_loss_scale=True)
534
535
536
537
538
539
540
541
542
543
    ]
    def_flags = {}
    def_flags['skip_eval'] = True
    def_flags['dtype'] = 'fp16'
    def_flags['enable_xla'] = True
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
    def_flags['train_steps'] = 600
    def_flags['log_steps'] = 100
    def_flags['distribution_strategy'] = 'default'

544
    super(TrivialKerasBenchmarkReal, self).__init__(
545
546
547
548
549
550
551
552
553
        output_dir=output_dir,
        flag_methods=flag_methods,
        default_flags=def_flags)

  def _run_and_report_benchmark(self):
    start_time_sec = time.time()
    stats = keras_imagenet_main.run(FLAGS)
    wall_time_sec = time.time() - start_time_sec

554
    super(TrivialKerasBenchmarkReal, self)._report_benchmark(
555
556
557
558
559
        stats,
        wall_time_sec,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)

560
561
562
563
564
565
566
567
568
569
570
  def benchmark_8_gpu_warmup(self):
    """Dummy test that runs over an epoch to warmup the machine."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_warmup')
    FLAGS.batch_size = 256
    FLAGS.train_steps = 700
    self._run_and_report_benchmark()

571
572
573
574
575
576
577
578
579
580
581
582
583
584
  def benchmark_1_gpu(self):
    """Test trivial Keras model (input pipeline) with 1 GPU."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

  def benchmark_graph_1_gpu(self):
    """Test trivial Keras model (input pipeline) with 1 GPU."""
    self._setup()

585
    FLAGS.num_gpus = 1
586
587
588
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
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
    FLAGS.enable_eager = False
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

  def benchmark_8_gpu(self):
    """Test trivial Keras model (input pipeline) with 8 GPUs."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
    FLAGS.batch_size = 256 * 8
    self._run_and_report_benchmark()

  def benchmark_8_gpu_tweaked(self):
    """Test trivial Keras model (input pipeline) with manual config tuning and
       8 GPUs.
    """
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_tweaked')
    FLAGS.batch_size = 256 * 8
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

  def benchmark_graph_8_gpu(self):
    """Test trivial Keras model (input pipeline) in legacy graph mode with 8
       GPUs.
    """
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
    FLAGS.batch_size = 256 * 8
    self._run_and_report_benchmark()

  def benchmark_graph_8_gpu_tweaked(self):
    """Test trivial Keras model (input pipeline) in legacy graph mode with
       manual config tuning and 8 GPUs.
    """
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu_tweaked')
    FLAGS.batch_size = 256 * 8
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

  def fill_report_object(self, stats):
640
    super(TrivialKerasBenchmarkReal, self).fill_report_object(
641
642
643
        stats,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)