keras_imagenet_benchmark.py 17.4 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
49
50
    flag_methods = [
        keras_common.define_keras_flags, imagenet_main.define_imagenet_flags
    ]
Toby Boyd's avatar
Toby Boyd committed
51

52
    self.data_dir = os.path.join(root_data_dir, 'imagenet')
53
54
    super(Resnet50KerasAccuracy, self).__init__(
        output_dir=output_dir, flag_methods=flag_methods)
55

Toby Boyd's avatar
Toby Boyd committed
56
  def benchmark_graph_8_gpu(self):
57
58
    """Test Keras model with Keras fit/dist_strat and 8 GPUs."""
    self._setup()
Toby Boyd's avatar
Toby Boyd committed
59
    FLAGS.num_gpus = 8
60
    FLAGS.data_dir = self.data_dir
61
    FLAGS.batch_size = 128 * 8
Toby Boyd's avatar
Toby Boyd committed
62
    FLAGS.train_epochs = 90
63
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
64
    FLAGS.dtype = 'fp32'
65
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
66
67

  def benchmark_8_gpu(self):
68
69
    """Test Keras model with eager, dist_strat and 8 GPUs."""
    self._setup()
Toby Boyd's avatar
Toby Boyd committed
70
    FLAGS.num_gpus = 8
71
    FLAGS.data_dir = self.data_dir
72
    FLAGS.batch_size = 128 * 8
Toby Boyd's avatar
Toby Boyd committed
73
    FLAGS.train_epochs = 90
74
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
75
76
    FLAGS.dtype = 'fp32'
    FLAGS.enable_eager = True
77
78
    # Add some thread tunings to improve performance.
    FLAGS.datasets_num_private_threads = 14
79
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
80

Reed's avatar
Reed committed
81
82
83
84
85
86
87
88
89
90
  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
91
92
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
Reed's avatar
Reed committed
93
94
95
96
97
98
99
100
101
102
103
104
105
    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
106
107
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
Reed's avatar
Reed committed
108
109
    self._run_and_report_benchmark()

110
111
112
113
114
115
  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
116
        stats,
117
        wall_time_sec,
Toby Boyd's avatar
Toby Boyd committed
118
119
        top_1_min=MIN_TOP_1_ACCURACY,
        top_1_max=MAX_TOP_1_ACCURACY,
120
        total_batch_size=FLAGS.batch_size,
Toby Boyd's avatar
Toby Boyd committed
121
        log_steps=100)
122
123
124
125

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

Toby Boyd's avatar
Toby Boyd committed
126
127
128
129
130

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

  def __init__(self, output_dir=None, default_flags=None):
131
132
133
    flag_methods = [
        keras_common.define_keras_flags, imagenet_main.define_imagenet_flags
    ]
Toby Boyd's avatar
Toby Boyd committed
134
135
136
137
138
139

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

140
141
  def _run_and_report_benchmark(self):
    start_time_sec = time.time()
Toby Boyd's avatar
Toby Boyd committed
142
    stats = keras_imagenet_main.run(FLAGS)
143
144
145
146
147
148
149
    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
150
151

  def benchmark_1_gpu_no_dist_strat(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
152
    """Test Keras model with 1 GPU, no distribution strategy."""
Toby Boyd's avatar
Toby Boyd committed
153
154
155
156
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
157
    FLAGS.distribution_strategy = 'off'
158
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
Toby Boyd's avatar
Toby Boyd committed
159
    FLAGS.batch_size = 128
160
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
161
162

  def benchmark_graph_1_gpu_no_dist_strat(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
163
    """Test Keras model in legacy graph mode with 1 GPU, no dist strat."""
Toby Boyd's avatar
Toby Boyd committed
164
165
166
167
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
168
    FLAGS.distribution_strategy = 'off'
169
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
Toby Boyd's avatar
Toby Boyd committed
170
    FLAGS.batch_size = 128
171
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
172
173

  def benchmark_1_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
174
    """Test Keras model with 1 GPU."""
Toby Boyd's avatar
Toby Boyd committed
175
176
177
178
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
179
    FLAGS.distribution_strategy = 'default'
180
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
Toby Boyd's avatar
Toby Boyd committed
181
    FLAGS.batch_size = 128
182
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
183

Haoyu Zhang's avatar
Haoyu Zhang committed
184
185
186
187
188
189
190
191
192
193
194
195
  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
196
  def benchmark_1_gpu_fp16(self):
197
    """Test Keras model with 1 GPU and fp16."""
Reed's avatar
Reed committed
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
    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()

  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()

Toby Boyd's avatar
Toby Boyd committed
221
  def benchmark_graph_1_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
222
    """Test Keras model in legacy graph mode with 1 GPU."""
Toby Boyd's avatar
Toby Boyd committed
223
224
225
226
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
227
    FLAGS.distribution_strategy = 'default'
228
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
Toby Boyd's avatar
Toby Boyd committed
229
    FLAGS.batch_size = 128
230
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
231

Haoyu Zhang's avatar
Haoyu Zhang committed
232
233
234
235
236
237
238
239
240
241
242
243
  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
244
  def benchmark_8_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
245
    """Test Keras model with 8 GPUs."""
Toby Boyd's avatar
Toby Boyd committed
246
247
248
249
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
250
    FLAGS.distribution_strategy = 'default'
251
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
252
    FLAGS.batch_size = 128 * 8  # 8 GPUs
253
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
254

255
  def benchmark_8_gpu_tweaked(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
256
    """Test Keras model with manual config tuning and 8 GPUs."""
257
258
259
260
261
262
263
264
265
266
    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
267
268
269
270
271
272
273
274
275
  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')
276
    FLAGS.batch_size = 128 * 8  # 8 GPUs
Haoyu Zhang's avatar
Haoyu Zhang committed
277
278
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
279
  def benchmark_8_gpu_fp16(self):
280
    """Test Keras model with 8 GPUs and fp16."""
Reed's avatar
Reed committed
281
282
283
    self._setup()

    FLAGS.num_gpus = 8
284
    FLAGS.dtype = 'fp16'
Reed's avatar
Reed committed
285
286
287
288
289
290
    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()

291
292
293
294
295
296
297
298
299
300
301
302
303
  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()

Reed's avatar
Reed committed
304
  def benchmark_xla_8_gpu_fp16(self):
305
    """Test Keras model with XLA, 8 GPUs and fp16."""
Reed's avatar
Reed committed
306
307
308
    self._setup()

    FLAGS.num_gpus = 8
309
    FLAGS.dtype = 'fp16'
Reed's avatar
Reed committed
310
311
312
313
314
315
316
    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()

317
318
319
320
321
322
323
324
325
326
327
328
329
330
  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()

331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
  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
347
  def benchmark_graph_8_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
348
    """Test Keras model in legacy graph mode with 8 GPUs."""
Toby Boyd's avatar
Toby Boyd committed
349
350
351
352
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
353
    FLAGS.distribution_strategy = 'default'
354
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
355
    FLAGS.batch_size = 128 * 8  # 8 GPUs
356
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
357

Haoyu Zhang's avatar
Haoyu Zhang committed
358
359
360
361
362
363
364
365
366
  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')
367
    FLAGS.batch_size = 128 * 8  # 8 GPUs
Haoyu Zhang's avatar
Haoyu Zhang committed
368
369
    self._run_and_report_benchmark()

370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
  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()

Toby Boyd's avatar
Toby Boyd committed
400
401
402
403
404
405
  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
406
407
408
409

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

410
  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
Toby Boyd's avatar
Toby Boyd committed
411
412
413
414
415
416
    def_flags = {}
    def_flags['skip_eval'] = True
    def_flags['use_synthetic_data'] = True
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

417
418
    super(Resnet50KerasBenchmarkSynth, self).__init__(
        output_dir=output_dir, default_flags=def_flags)
Toby Boyd's avatar
Toby Boyd committed
419
420
421
422
423


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

424
  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
Toby Boyd's avatar
Toby Boyd committed
425
426
    def_flags = {}
    def_flags['skip_eval'] = True
427
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
Toby Boyd's avatar
Toby Boyd committed
428
429
430
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

431
432
    super(Resnet50KerasBenchmarkReal, self).__init__(
        output_dir=output_dir, default_flags=def_flags)
433
434


435
class TrivialKerasBenchmarkReal(keras_benchmark.KerasBenchmark):
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
  """Trivial model with real data benchmark tests."""

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
    flag_methods = [
        keras_common.define_keras_flags, imagenet_main.define_imagenet_flags
    ]
    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'

451
    super(TrivialKerasBenchmarkReal, self).__init__(
452
453
454
455
456
457
458
459
460
        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

461
    super(TrivialKerasBenchmarkReal, self)._report_benchmark(
462
463
464
465
466
        stats,
        wall_time_sec,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)

467
468
469
470
471
472
473
474
475
476
477
  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()

478
479
480
481
482
483
484
485
486
487
488
489
490
491
  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()

492
    FLAGS.num_gpus = 1
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
    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):
547
    super(TrivialKerasBenchmarkReal, self).fill_report_object(
548
549
550
        stats,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)