keras_imagenet_benchmark.py 39.7 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

from absl import flags
22
import tensorflow as tf  # pylint: disable=g-bad-import-order
23

24
from official.benchmark import keras_benchmark
25
from official.utils.testing import benchmark_wrappers
26
from official.vision.image_classification import resnet_imagenet_main
27

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 = [resnet_imagenet_main.define_imagenet_keras_flags]
Toby Boyd's avatar
Toby Boyd committed
49

50
    self.data_dir = os.path.join(root_data_dir, 'imagenet')
51
52
    super(Resnet50KerasAccuracy, self).__init__(
        output_dir=output_dir, flag_methods=flag_methods)
53

Toby Boyd's avatar
Toby Boyd committed
54
  def benchmark_graph_8_gpu(self):
55
56
    """Test Keras model with Keras fit/dist_strat and 8 GPUs."""
    self._setup()
Toby Boyd's avatar
Toby Boyd committed
57
    FLAGS.num_gpus = 8
58
    FLAGS.data_dir = self.data_dir
59
    FLAGS.batch_size = 128 * 8
Toby Boyd's avatar
Toby Boyd committed
60
    FLAGS.train_epochs = 90
61
    FLAGS.epochs_between_evals = 10
62
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
63
    FLAGS.dtype = 'fp32'
64
    FLAGS.use_tensor_lr = True
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.epochs_between_evals = 10
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
    FLAGS.use_tensor_lr = True
81
    self._run_and_report_benchmark()
82

83
84
85
86
87
88
89
90
91
  def benchmark_8_gpu_amp(self):
    """Test Keras model with eager, dist_strat and 8 GPUs with automatic mixed precision."""
    self._setup()
    FLAGS.num_gpus = 8
    FLAGS.data_dir = self.data_dir
    FLAGS.batch_size = 128 * 8
    FLAGS.train_epochs = 90
    FLAGS.epochs_between_evals = 10
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp')
Vinh Nguyen's avatar
Vinh Nguyen committed
92
    FLAGS.dtype = 'fp16'
93
    FLAGS.enable_eager = True
94
    FLAGS.fp16_implementation = 'graph_rewrite'
95
96
97
98
    # Add some thread tunings to improve performance.
    FLAGS.datasets_num_private_threads = 14
    FLAGS.use_tensor_lr = True
    self._run_and_report_benchmark()
99

Reed's avatar
Reed committed
100
101
102
103
104
105
106
  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
107
    FLAGS.epochs_between_evals = 10
Reed's avatar
Reed committed
108
109
110
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
111
112
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
113
    FLAGS.use_tensor_lr = True
Reed's avatar
Reed committed
114
115
116
117
118
119
120
121
122
    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
123
    FLAGS.epochs_between_evals = 10
Reed's avatar
Reed committed
124
125
126
127
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
128
129
    # Thread tuning to improve performance.
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
130
    FLAGS.use_tensor_lr = True
Reed's avatar
Reed committed
131
132
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
  def benchmark_8_gpu_mlperf_like(self):
    """Test similar to the rules for MLPerf 0.5.

    Listed below are reasons this comparison is not to the MLSpec, but this is
    still a decent directional measurement:
      - Eval is every 4 epochs and again at the end. ~2 extra times.
      - Learning rate is not tuned to hit 75%, but we know the model is correct.
      - We measure total time and MLPerf 0.5 excluded some startup time.
      - Eval is not on the total set, need to set eval batch_size where
        8*batch_size/50K is even. 250 is a good number.
      - Not sure if we are doing any extra or too few steps due to epoch bleed.
    """
    self._setup()
    FLAGS.num_gpus = 8
    FLAGS.data_dir = self.data_dir
    FLAGS.batch_size = 256 * 8
    FLAGS.train_epochs = 61
    FLAGS.epochs_between_evals = 4
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mlperf_like')
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
155
    self._run_and_report_benchmark(top_1_min=0.736)
Toby Boyd's avatar
Toby Boyd committed
156

157
158
159
160
161
162
163
  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
164
    FLAGS.epochs_between_evals = 10
165
166
167
168
169
170
171
    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'
172
    FLAGS.use_tensor_lr = True
173
    self._run_and_report_benchmark(top_1_min=0.736)
174

175
  @benchmark_wrappers.enable_runtime_flags
176
177
178
  def _run_and_report_benchmark(self,
                                top_1_min=MIN_TOP_1_ACCURACY,
                                top_1_max=MAX_TOP_1_ACCURACY):
179
    start_time_sec = time.time()
180
    stats = resnet_imagenet_main.run(flags.FLAGS)
181
182
183
    wall_time_sec = time.time() - start_time_sec

    super(Resnet50KerasAccuracy, self)._report_benchmark(
Toby Boyd's avatar
Toby Boyd committed
184
        stats,
185
        wall_time_sec,
186
187
        top_1_min=top_1_min,
        top_1_max=top_1_max,
188
        total_batch_size=FLAGS.batch_size,
Toby Boyd's avatar
Toby Boyd committed
189
        log_steps=100)
190
191
192
193

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

Toby Boyd's avatar
Toby Boyd committed
194
195
196
197

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

David Chen's avatar
David Chen committed
198
  def __init__(self, output_dir=None, default_flags=None, tpu=None):
199
    flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
Toby Boyd's avatar
Toby Boyd committed
200
201
202
203

    super(Resnet50KerasBenchmarkBase, self).__init__(
        output_dir=output_dir,
        flag_methods=flag_methods,
David Chen's avatar
David Chen committed
204
205
        default_flags=default_flags,
        tpu=tpu)
Toby Boyd's avatar
Toby Boyd committed
206

207
  @benchmark_wrappers.enable_runtime_flags
208
  def _run_and_report_benchmark(self, skip_steps=None):
209
    start_time_sec = time.time()
210
    stats = resnet_imagenet_main.run(FLAGS)
211
    wall_time_sec = time.time() - start_time_sec
212
    # Number of logged step time entries that are excluded in performance
213
214
215
    # report. We keep results from last 100 batches, or skip the steps based on
    # input skip_steps.
    warmup = (skip_steps or (FLAGS.train_steps - 100)) // FLAGS.log_steps
216
217
218
219
220

    super(Resnet50KerasBenchmarkBase, self)._report_benchmark(
        stats,
        wall_time_sec,
        total_batch_size=FLAGS.batch_size,
221
        log_steps=FLAGS.log_steps,
David Chen's avatar
David Chen committed
222
223
        warmup=warmup,
        start_time_sec=start_time_sec)
Toby Boyd's avatar
Toby Boyd committed
224
225

  def benchmark_1_gpu_no_dist_strat(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
226
    """Test Keras model with 1 GPU, no distribution strategy."""
Toby Boyd's avatar
Toby Boyd committed
227
228
229
230
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
231
    FLAGS.distribution_strategy = 'off'
232
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_no_dist_strat')
Toby Boyd's avatar
Toby Boyd committed
233
    FLAGS.batch_size = 128
234
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
235

236
237
238
239
240
241
242
243
244
245
246
247
248
  def benchmark_1_gpu_no_dist_strat_run_eagerly(self):
    """Test Keras model with 1 GPU, no distribution strategy, run eagerly."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.run_eagerly = True
    FLAGS.distribution_strategy = 'off'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_1_gpu_no_dist_strat_run_eagerly')
    FLAGS.batch_size = 64
    self._run_and_report_benchmark()

249
250
251
252
253
254
255
256
257
258
259
260
261
262
  def benchmark_1_gpu_no_dist_strat_run_eagerly_tweaked(self):
    """Test Keras model with 1 GPU, no distribution strategy, run eagerly."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.run_eagerly = True
    FLAGS.explicit_gpu_placement = True
    FLAGS.distribution_strategy = 'off'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_1_gpu_no_dist_strat_run_eagerly_tweaked')
    FLAGS.batch_size = 64
    self._run_and_report_benchmark()

263
264
265
266
267
268
269
270
271
272
273
274
275
276
  def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16(self):
    """Test with 1 GPU, no distribution strategy, fp16, run eagerly."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.run_eagerly = True
    FLAGS.distribution_strategy = 'off'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_1_gpu_no_dist_strat_run_eagerly_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 128
    self._run_and_report_benchmark()

277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
  def benchmark_1_gpu_no_dist_strat_run_eagerly_fp16_tweaked(self):
    """Test with 1 GPU, no distribution strategy, fp16, run eagerly."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.run_eagerly = True
    FLAGS.explicit_gpu_placement = True
    FLAGS.distribution_strategy = 'off'
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_1_gpu_no_dist_strat_run_eagerly_fp16_tweaked')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 128
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
292
  def benchmark_graph_1_gpu_no_dist_strat(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
293
    """Test Keras model in legacy graph mode with 1 GPU, no dist strat."""
Toby Boyd's avatar
Toby Boyd committed
294
295
296
297
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
298
    FLAGS.distribution_strategy = 'off'
299
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_no_dist_strat')
300
    FLAGS.batch_size = 96  # BatchNorm is less efficient in legacy graph mode
301
    # due to its reliance on v1 cond.
302
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
303
304

  def benchmark_1_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
305
    """Test Keras model with 1 GPU."""
Toby Boyd's avatar
Toby Boyd committed
306
307
308
309
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
310
    FLAGS.distribution_strategy = 'one_device'
311
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu')
Toby Boyd's avatar
Toby Boyd committed
312
    FLAGS.batch_size = 128
313
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
314

315
316
317
318
319
320
  def benchmark_1_gpu_amp(self):
    """Test Keras model with 1 GPU with automatic mixed precision."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
Vinh Nguyen's avatar
Vinh Nguyen committed
321
    FLAGS.dtype = 'fp16'
322
    FLAGS.fp16_implementation = 'graph_rewrite'
323
    FLAGS.distribution_strategy = 'one_device'
324
325
326
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_amp')
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()
327

Haoyu Zhang's avatar
Haoyu Zhang committed
328
329
330
331
332
333
334
  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
335
    FLAGS.distribution_strategy = 'one_device'
Haoyu Zhang's avatar
Haoyu Zhang committed
336
337
338
339
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu')
    FLAGS.batch_size = 128
    self._run_and_report_benchmark()

340
341
342
343
344
345
  def benchmark_xla_1_gpu_amp(self):
    """Test Keras model with XLA and 1 GPU with automatic mixed precision."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
Vinh Nguyen's avatar
Vinh Nguyen committed
346
    FLAGS.dtype = 'fp16'
347
    FLAGS.fp16_implementation = 'graph_rewrite'
348
    FLAGS.enable_xla = True
349
    FLAGS.distribution_strategy = 'one_device'
350
351
352
353
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_amp')
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
354
  def benchmark_1_gpu_fp16(self):
355
    """Test Keras model with 1 GPU and fp16."""
Reed's avatar
Reed committed
356
357
358
359
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
360
    FLAGS.distribution_strategy = 'one_device'
Reed's avatar
Reed committed
361
362
363
364
365
    FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

366
367
368
369
370
371
  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
372
    FLAGS.distribution_strategy = 'one_device'
373
374
375
376
377
378
    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
379
380
381
382
383
384
385
  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
386
    FLAGS.distribution_strategy = 'one_device'
Reed's avatar
Reed committed
387
388
389
390
391
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

392
393
394
395
396
397
398
  def benchmark_xla_1_gpu_fp16_tweaked(self):
    """Test Keras model with XLA, 1 GPU, fp16, and manual config tuning."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
399
    FLAGS.distribution_strategy = 'one_device'
400
401
402
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_1_gpu_fp16_tweaked')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
403
    FLAGS.use_tensor_lr = True
404
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
405
406
    self._run_and_report_benchmark()

407
408
409
410
411
412
413
  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
414
    FLAGS.distribution_strategy = 'one_device'
415
416
417
418
419
420
    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
421
  def benchmark_graph_1_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
422
    """Test Keras model in legacy graph mode with 1 GPU."""
Toby Boyd's avatar
Toby Boyd committed
423
424
425
426
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
427
    FLAGS.distribution_strategy = 'one_device'
428
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu')
Toby Boyd's avatar
Toby Boyd committed
429
    FLAGS.batch_size = 128
430
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
431

Haoyu Zhang's avatar
Haoyu Zhang committed
432
433
434
435
436
437
438
  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
439
    FLAGS.distribution_strategy = 'one_device'
Haoyu Zhang's avatar
Haoyu Zhang committed
440
441
442
443
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu')
    FLAGS.batch_size = 128
    self._run_and_report_benchmark()

444
445
446
447
448
  def benchmark_graph_1_gpu_fp16(self):
    """Test Keras model in legacy graph mode with 1 GPU and fp16."""
    self._setup()

    FLAGS.num_gpus = 1
449
    FLAGS.dtype = 'fp16'
450
    FLAGS.enable_eager = False
451
    FLAGS.distribution_strategy = 'one_device'
452
453
454
455
456
457
458
459
460
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_1_gpu_fp16')
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

  def benchmark_graph_xla_1_gpu_fp16(self):
    """Test Keras model in legacy graph mode with 1 GPU, fp16 and XLA."""
    self._setup()

    FLAGS.num_gpus = 1
461
    FLAGS.dtype = 'fp16'
462
463
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
464
    FLAGS.distribution_strategy = 'one_device'
465
466
467
468
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_1_gpu_fp16')
    FLAGS.batch_size = 256
    self._run_and_report_benchmark()

469
  def benchmark_graph_xla_1_gpu_fp16_tweaked(self):
470
    """Test Keras model in legacy graph with 1 GPU, fp16, XLA, and tuning."""
471
472
473
474
475
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
476
    FLAGS.distribution_strategy = 'one_device'
477
478
479
480
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_xla_1_gpu_fp16_tweaked')
    FLAGS.dtype = 'fp16'
    FLAGS.batch_size = 256
481
    FLAGS.use_tensor_lr = True
482
483
484
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
485
  def benchmark_8_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
486
    """Test Keras model with 8 GPUs."""
Toby Boyd's avatar
Toby Boyd committed
487
488
489
490
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
491
    FLAGS.distribution_strategy = 'mirrored'
492
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
493
    FLAGS.batch_size = 128 * 8  # 8 GPUs
494
    self._run_and_report_benchmark()
495

496
497
498
499
500
501
  def benchmark_8_gpu_amp(self):
    """Test Keras model with 8 GPUs with automatic mixed precision."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
Vinh Nguyen's avatar
Vinh Nguyen committed
502
    FLAGS.dtype = 'fp16'
503
    FLAGS.fp16_implementation = 'graph_rewrite'
504
    FLAGS.distribution_strategy = 'mirrored'
505
506
507
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()
508

509
  def benchmark_8_gpu_tweaked(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
510
    """Test Keras model with manual config tuning and 8 GPUs."""
511
512
513
514
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
515
    FLAGS.distribution_strategy = 'mirrored'
516
517
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_tweaked')
    FLAGS.batch_size = 128 * 8  # 8 GPUs
518
    FLAGS.use_tensor_lr = True
519
    FLAGS.datasets_num_private_threads = 14
520
521
    self._run_and_report_benchmark()

Haoyu Zhang's avatar
Haoyu Zhang committed
522
523
524
525
526
527
528
  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
529
    FLAGS.distribution_strategy = 'mirrored'
Haoyu Zhang's avatar
Haoyu Zhang committed
530
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu')
531
    FLAGS.batch_size = 128 * 8  # 8 GPUs
Haoyu Zhang's avatar
Haoyu Zhang committed
532
533
    self._run_and_report_benchmark()

534
535
536
537
538
539
  def benchmark_xla_8_gpu_amp(self):
    """Test Keras model with XLA and 8 GPUs with automatic mixed precision."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
Vinh Nguyen's avatar
Vinh Nguyen committed
540
    FLAGS.dtype = 'fp16'
541
    FLAGS.fp16_implementation = 'graph_rewrite'
542
    FLAGS.enable_xla = True
543
    FLAGS.distribution_strategy = 'mirrored'
544
545
546
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_amp')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()
547

548
549
550
551
552
553
554
  def benchmark_xla_8_gpu_tweaked(self):
    """Test Keras model with manual config tuning, 8 GPUs, and XLA."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
555
    FLAGS.distribution_strategy = 'mirrored'
556
557
558
559
560
561
562
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_tweaked')
    FLAGS.batch_size = 128 * 8
    FLAGS.use_tensor_lr = True
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    FLAGS.datasets_num_private_threads = 24
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
563
  def benchmark_8_gpu_fp16(self):
564
    """Test Keras model with 8 GPUs and fp16."""
Reed's avatar
Reed committed
565
566
567
    self._setup()

    FLAGS.num_gpus = 8
568
    FLAGS.dtype = 'fp16'
Reed's avatar
Reed committed
569
    FLAGS.enable_eager = True
570
    FLAGS.distribution_strategy = 'mirrored'
Reed's avatar
Reed committed
571
572
573
574
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()

575
  def benchmark_8_gpu_fp16_tweaked(self):
576
    """Test Keras model with 8 GPUs, fp16, and manual config tuning."""
577
578
579
580
581
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
582
    FLAGS.distribution_strategy = 'mirrored'
583
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_fp16_tweaked')
584
    FLAGS.batch_size = 256 * 8  # 8 GPUs
585
    FLAGS.use_tensor_lr = True
586
587
588
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

589
  def benchmark_8_gpu_fp16_dynamic_tweaked(self):
Toby Boyd's avatar
Toby Boyd committed
590
    """Test Keras model with 8 GPUs, fp16, dynamic loss scaling, and tuned."""
591
592
593
594
595
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
596
    FLAGS.distribution_strategy = 'mirrored'
597
598
599
600
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_8_gpu_fp16_dynamic_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.loss_scale = 'dynamic'
601
    FLAGS.use_tensor_lr = True
602
603
604
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

Reed's avatar
Reed committed
605
  def benchmark_xla_8_gpu_fp16(self):
606
    """Test Keras model with XLA, 8 GPUs and fp16."""
Reed's avatar
Reed committed
607
608
609
    self._setup()

    FLAGS.num_gpus = 8
610
    FLAGS.dtype = 'fp16'
Reed's avatar
Reed committed
611
612
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
613
    FLAGS.distribution_strategy = 'mirrored'
Reed's avatar
Reed committed
614
615
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
616
617
    self._run_and_report_benchmark()

618
619
620
621
622
623
624
625
  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
626
    FLAGS.distribution_strategy = 'mirrored'
627
628
    FLAGS.model_dir = self._get_model_dir('benchmark_xla_8_gpu_fp16_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
629
    FLAGS.use_tensor_lr = True
630
631
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    FLAGS.datasets_num_private_threads = 48
632
633
    self._run_and_report_benchmark()

634
  def benchmark_xla_8_gpu_fp16_tweaked_delay_measure(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
635
636
637
    """Test with manual config tuning, XLA, 8 GPUs and fp16.

    Delay performance measurement for stable performance on 96 vCPU platforms.
638
639
640
641
642
643
644
    """
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = True
    FLAGS.enable_xla = True
645
    FLAGS.distribution_strategy = 'mirrored'
646
647
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_xla_8_gpu_fp16_tweaked_delay_measure')
648
    FLAGS.batch_size = 256 * 8
649
650
651
652
653
    FLAGS.use_tensor_lr = True
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    FLAGS.train_steps = 310
    self._run_and_report_benchmark()

654
655
656
657
658
659
660
661
  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
662
    FLAGS.distribution_strategy = 'mirrored'
663
664
665
666
    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'
667
    FLAGS.use_tensor_lr = True
668
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
669
    FLAGS.datasets_num_private_threads = 48
670
671
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
672
  def benchmark_graph_8_gpu(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
673
    """Test Keras model in legacy graph mode with 8 GPUs."""
Toby Boyd's avatar
Toby Boyd committed
674
675
676
677
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
678
    FLAGS.distribution_strategy = 'mirrored'
679
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu')
Toby Boyd's avatar
Toby Boyd committed
680
    FLAGS.batch_size = 128 * 8  # 8 GPUs
681
    self._run_and_report_benchmark()
Toby Boyd's avatar
Toby Boyd committed
682

Haoyu Zhang's avatar
Haoyu Zhang committed
683
684
685
686
687
688
689
  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
690
    FLAGS.distribution_strategy = 'mirrored'
Haoyu Zhang's avatar
Haoyu Zhang committed
691
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_xla_8_gpu')
692
    FLAGS.batch_size = 128 * 8  # 8 GPUs
Haoyu Zhang's avatar
Haoyu Zhang committed
693
694
    self._run_and_report_benchmark()

695
696
697
698
699
700
701
  def benchmark_graph_8_gpu_fp16(self):
    """Test Keras model in legacy graph mode with 8 GPUs and fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
702
    FLAGS.distribution_strategy = 'mirrored'
703
704
705
706
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu_fp16')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    self._run_and_report_benchmark()

707
708
709
710
711
712
713
714
  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
715
    FLAGS.distribution_strategy = 'mirrored'
716
717
718
719
    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()

720
  def benchmark_graph_8_gpu_fp16_tweaked(self):
721
    """Test Keras model in legacy graph mode, tuning, 8 GPUs, and FP16."""
722
723
724
725
726
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
727
    FLAGS.distribution_strategy = 'mirrored'
728
729
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu_fp16_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
730
    FLAGS.use_tensor_lr = True
731
732
733
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

734
  def benchmark_graph_xla_8_gpu_fp16_tweaked(self):
735
    """Test Keras model in legacy graph tuning, XLA_FP16, 8 GPUs and fp16."""
736
737
738
739
740
741
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
742
    FLAGS.distribution_strategy = 'mirrored'
743
744
745
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_xla_8_gpu_fp16_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
746
    FLAGS.use_tensor_lr = True
747
748
749
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

750
  def benchmark_graph_xla_8_gpu_fp16_tweaked_delay_measure(self):
Haoyu Zhang's avatar
Haoyu Zhang committed
751
752
753
    """Test in legacy graph mode with manual config tuning, XLA, 8 GPUs, fp16.

    Delay performance measurement for stable performance on 96 vCPU platforms.
754
755
756
757
758
759
760
    """
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
    FLAGS.enable_xla = True
761
    FLAGS.distribution_strategy = 'mirrored'
762
763
764
765
766
767
768
769
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_xla_8_gpu_fp16_tweaked_delay_measure')
    FLAGS.batch_size = 256 * 8
    FLAGS.use_tensor_lr = True
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    FLAGS.train_steps = 310
    self._run_and_report_benchmark()

770
771
772
773
774
775
776
  def benchmark_graph_8_gpu_fp16_dynamic_tweaked(self):
    """Test graph Keras with config tuning, 8 GPUs and dynamic fp16."""
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = False
777
    FLAGS.distribution_strategy = 'mirrored'
778
779
780
781
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_8_gpu_fp16_dynamic_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
    FLAGS.loss_scale = 'dynamic'
782
    FLAGS.use_tensor_lr = True
783
784
785
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

786
787
788
789
790
791
792
793
  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
794
    FLAGS.distribution_strategy = 'mirrored'
795
796
797
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_graph_xla_8_gpu_fp16_dynamic_tweaked')
    FLAGS.batch_size = 256 * 8  # 8 GPUs
798
    FLAGS.use_tensor_lr = True
799
800
801
802
    FLAGS.loss_scale = 'dynamic'
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
    self._run_and_report_benchmark()

David Chen's avatar
David Chen committed
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
  def benchmark_2x2_tpu_fp16(self):
    """Test Keras model with 2x2 TPU, fp16."""
    self._setup()

    FLAGS.dtype = 'bf16'
    FLAGS.distribution_strategy = 'tpu'
    FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_fp16')
    FLAGS.batch_size = 1024
    self._run_and_report_benchmark()

  def benchmark_4x4_tpu_fp16(self):
    """Test Keras model with 4x4 TPU, fp16."""
    self._setup()

    FLAGS.dtype = 'bf16'
    FLAGS.distribution_strategy = 'tpu'
    FLAGS.model_dir = self._get_model_dir('benchmark_4x4_tpu_fp16')
    FLAGS.batch_size = 4096
    self._run_and_report_benchmark()

Toby Boyd's avatar
Toby Boyd committed
823
824
825
826
827
828
  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
829
830
831
832

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

David Chen's avatar
David Chen committed
833
  def __init__(self, output_dir=None, root_data_dir=None, tpu=None, **kwargs):
Toby Boyd's avatar
Toby Boyd committed
834
835
    def_flags = {}
    def_flags['skip_eval'] = True
836
    def_flags['report_accuracy_metrics'] = False
Toby Boyd's avatar
Toby Boyd committed
837
838
839
840
    def_flags['use_synthetic_data'] = True
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

841
    super(Resnet50KerasBenchmarkSynth, self).__init__(
David Chen's avatar
David Chen committed
842
        output_dir=output_dir, default_flags=def_flags, tpu=tpu)
Toby Boyd's avatar
Toby Boyd committed
843
844
845
846
847


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

David Chen's avatar
David Chen committed
848
  def __init__(self, output_dir=None, root_data_dir=None, tpu=None, **kwargs):
Toby Boyd's avatar
Toby Boyd committed
849
850
    def_flags = {}
    def_flags['skip_eval'] = True
851
    def_flags['report_accuracy_metrics'] = False
852
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
Toby Boyd's avatar
Toby Boyd committed
853
854
855
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

856
    super(Resnet50KerasBenchmarkReal, self).__init__(
David Chen's avatar
David Chen committed
857
        output_dir=output_dir, default_flags=def_flags, tpu=tpu)
858
859


860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
class Resnet50KerasBenchmarkRemoteData(Resnet50KerasBenchmarkBase):
  """Resnet50 real data (stored in remote storage) benchmark tests."""

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
    def_flags = {}
    def_flags['skip_eval'] = True
    def_flags['report_accuracy_metrics'] = False
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
    # Defining multiple epochs overrides the train_steps setting in benchmarks.
    def_flags['train_epochs'] = 2
    # Cache dataset so performance is stable after the first epoch.
    def_flags['training_dataset_cache'] = True
    def_flags['log_steps'] = 100

    super(Resnet50KerasBenchmarkRemoteData, self).__init__(
        output_dir=output_dir, default_flags=def_flags)

877
  @benchmark_wrappers.enable_runtime_flags
878
879
880
881
882
883
  def _run_and_report_benchmark(self):
    # skip the first epoch for performance measurement.
    super(Resnet50KerasBenchmarkRemoteData,
          self)._run_and_report_benchmark(skip_steps=600)


884
class TrivialKerasBenchmarkReal(keras_benchmark.KerasBenchmark):
885
886
887
  """Trivial model with real data benchmark tests."""

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
888
    flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
Toby Boyd's avatar
Toby Boyd committed
889

890
    def_flags = {}
891
    def_flags['use_trivial_model'] = True
892
    def_flags['skip_eval'] = True
893
    def_flags['report_accuracy_metrics'] = False
894
    def_flags['use_tensor_lr'] = True
895
    def_flags['dtype'] = 'fp16'
896
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
897
898
    def_flags['train_steps'] = 600
    def_flags['log_steps'] = 100
899
    def_flags['distribution_strategy'] = 'mirrored'
900

901
    super(TrivialKerasBenchmarkReal, self).__init__(
902
903
904
905
        output_dir=output_dir,
        flag_methods=flag_methods,
        default_flags=def_flags)

906
  @benchmark_wrappers.enable_runtime_flags
907
908
  def _run_and_report_benchmark(self):
    start_time_sec = time.time()
909
    stats = resnet_imagenet_main.run(FLAGS)
910
911
    wall_time_sec = time.time() - start_time_sec

912
    super(TrivialKerasBenchmarkReal, self)._report_benchmark(
913
914
915
916
917
        stats,
        wall_time_sec,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)

918
919
920
921
922
923
924
  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')
925
    FLAGS.batch_size = 256 * 8
926
927
928
    FLAGS.train_steps = 700
    self._run_and_report_benchmark()

929
930
931
932
933
934
  def benchmark_1_gpu(self):
    """Test trivial Keras model (input pipeline) with 1 GPU."""
    self._setup()

    FLAGS.num_gpus = 1
    FLAGS.enable_eager = True
935
    FLAGS.enable_xla = True
936
937
938
939
940
941
942
943
    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()

944
    FLAGS.num_gpus = 1
945
    FLAGS.enable_eager = False
946
    FLAGS.enable_xla = True
947
948
949
950
951
952
953
954
955
956
    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
957
    FLAGS.enable_xla = True
958
959
960
961
962
    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):
963
    """Test trivial Keras model with tuning and 8 GPUs."""
964
965
966
967
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = True
968
    FLAGS.enable_xla = True
969
970
971
    FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_tweaked')
    FLAGS.batch_size = 256 * 8
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
972
    FLAGS.datasets_num_private_threads = 48
973
974
975
    self._run_and_report_benchmark()

  def benchmark_graph_8_gpu(self):
976
    """Test trivial Keras model in legacy graph mode with 8 GPUs."""
977
978
979
980
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
981
    FLAGS.enable_xla = True
982
983
984
985
986
    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):
987
    """Test trivial Keras model in legacy graph mode with tuning and 8 GPUs."""
988
989
990
991
    self._setup()

    FLAGS.num_gpus = 8
    FLAGS.enable_eager = False
992
    FLAGS.enable_xla = True
993
994
995
    FLAGS.model_dir = self._get_model_dir('benchmark_graph_8_gpu_tweaked')
    FLAGS.batch_size = 256 * 8
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
996
    FLAGS.datasets_num_private_threads = 48
997
998
999
    self._run_and_report_benchmark()

  def fill_report_object(self, stats):
1000
    super(TrivialKerasBenchmarkReal, self).fill_report_object(
1001
1002
1003
        stats,
        total_batch_size=FLAGS.batch_size,
        log_steps=FLAGS.log_steps)
1004
1005


1006
1007
1008
1009
1010
class Resnet50MultiWorkerKerasAccuracy(keras_benchmark.KerasBenchmark):
  """Resnet50 distributed accuracy tests with multiple workers."""

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
    flag_methods = [resnet_imagenet_main.define_imagenet_keras_flags]
1011
    self.data_dir = os.path.join(root_data_dir, 'imagenet')
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
    super(Resnet50MultiWorkerKerasAccuracy, self).__init__(
        output_dir=output_dir, flag_methods=flag_methods)

  def _benchmark_common(self, eager, num_workers, all_reduce_alg):
    """Common to all benchmarks in this class."""
    self._setup()

    num_gpus = 8
    FLAGS.num_gpus = num_gpus
    FLAGS.data_dir = self.data_dir
    FLAGS.train_epochs = 90
    FLAGS.epochs_between_evals = 10
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = eager
    FLAGS.enable_xla = False
    FLAGS.distribution_strategy = 'multi_worker_mirrored'
    FLAGS.use_tensor_lr = True
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
1030
    FLAGS.datasets_num_private_threads = 32
1031
1032
1033
1034
1035
1036
1037
1038
    FLAGS.model_dir = self._get_model_dir(
        'benchmark_{}_8_gpu_{}_worker_fp16_{}_tweaked'.format(
            'eager' if eager else 'graph', num_workers, all_reduce_alg))
    FLAGS.batch_size = 256 * num_gpus * num_workers
    FLAGS.all_reduce_alg = all_reduce_alg

    self._run_and_report_benchmark()

1039
  @benchmark_wrappers.enable_runtime_flags
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
  def _run_and_report_benchmark(self,
                                top_1_min=MIN_TOP_1_ACCURACY,
                                top_1_max=MAX_TOP_1_ACCURACY):
    start_time_sec = time.time()
    stats = resnet_imagenet_main.run(flags.FLAGS)
    wall_time_sec = time.time() - start_time_sec

    super(Resnet50MultiWorkerKerasAccuracy, self)._report_benchmark(
        stats,
        wall_time_sec,
        top_1_min=top_1_min,
        top_1_max=top_1_max,
        total_batch_size=FLAGS.batch_size,
        log_steps=100)

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

  def benchmark_eager_8_gpu_2_workers_fp16_ring_tweaked(self):
    """Eager, 8 GPUs per worker, 2 workers, fp16, ring all-reduce."""
    self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='ring')

  def benchmark_eager_8_gpu_2_workers_fp16_nccl_tweaked(self):
    """Eager, 8 GPUs per worker, 2 workers, fp16, nccl all-reduce."""
    self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='nccl')

  def benchmark_eager_8_gpu_8_workers_fp16_ring_tweaked(self):
    """Eager, 8 GPUs per worker, 8 workers, fp16, ring all-reduce."""
    self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='ring')

  def benchmark_eager_8_gpu_8_workers_fp16_nccl_tweaked(self):
    """Eager, 8 GPUs per worker, 8 workers, fp16, nccl all-reduce."""
    self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='nccl')


1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
class Resnet50MultiWorkerKerasBenchmark(Resnet50KerasBenchmarkBase):
  """Resnet50 distributed benchmark tests with multiple workers."""

  def __init__(self, output_dir=None, default_flags=None):
    super(Resnet50MultiWorkerKerasBenchmark, self).__init__(
        output_dir=output_dir, default_flags=default_flags)

  def _benchmark_common(self, eager, num_workers, all_reduce_alg):
    """Common to all benchmarks in this class."""
    self._setup()

    num_gpus = 8
    FLAGS.num_gpus = num_gpus
    FLAGS.dtype = 'fp16'
    FLAGS.enable_eager = eager
    FLAGS.enable_xla = False
    FLAGS.distribution_strategy = 'multi_worker_mirrored'
    FLAGS.use_tensor_lr = True
    FLAGS.tf_gpu_thread_mode = 'gpu_private'
1094
    FLAGS.datasets_num_private_threads = 32
1095
    FLAGS.model_dir = self._get_model_dir(
1096
1097
        'benchmark_{}_8_gpu_{}_worker_fp16_{}_tweaked'.format(
            'eager' if eager else 'graph', num_workers, all_reduce_alg))
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
    FLAGS.batch_size = 256 * num_gpus * num_workers
    FLAGS.all_reduce_alg = all_reduce_alg

    self._run_and_report_benchmark()

  def benchmark_eager_8_gpu_1_worker_fp16_ring_tweaked(self):
    """Eager, 8 GPUs per worker, 1 worker, fp16, ring all-reduce."""
    self._benchmark_common(eager=True, num_workers=1, all_reduce_alg='ring')

  def benchmark_eager_8_gpu_1_worker_fp16_nccl_tweaked(self):
    """Eager, 8 GPUs per worker, 1 worker, fp16, nccl all-reduce."""
    self._benchmark_common(eager=True, num_workers=1, all_reduce_alg='nccl')

  def benchmark_eager_8_gpu_2_workers_fp16_ring_tweaked(self):
    """Eager, 8 GPUs per worker, 2 workers, fp16, ring all-reduce."""
    self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='ring')

  def benchmark_eager_8_gpu_2_workers_fp16_nccl_tweaked(self):
    """Eager, 8 GPUs per worker, 2 workers, fp16, nccl all-reduce."""
    self._benchmark_common(eager=True, num_workers=2, all_reduce_alg='nccl')

  def benchmark_eager_8_gpu_8_workers_fp16_ring_tweaked(self):
    """Eager, 8 GPUs per worker, 8 workers, fp16, ring all-reduce."""
    self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='ring')

  def benchmark_eager_8_gpu_8_workers_fp16_nccl_tweaked(self):
    """Eager, 8 GPUs per worker, 8 workers, fp16, nccl all-reduce."""
    self._benchmark_common(eager=True, num_workers=8, all_reduce_alg='nccl')


Ayush Dubey's avatar
Ayush Dubey committed
1128
class Resnet50MultiWorkerKerasBenchmarkSynth(Resnet50MultiWorkerKerasBenchmark):
1129
  """Resnet50 multi-worker synthetic data benchmark tests."""
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
    def_flags = {}
    def_flags['skip_eval'] = True
    def_flags['report_accuracy_metrics'] = False
    def_flags['use_synthetic_data'] = True
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

    super(Resnet50MultiWorkerKerasBenchmarkSynth, self).__init__(
        output_dir=output_dir, default_flags=def_flags)


1143
1144
1145
1146
1147
1148
1149
class Resnet50MultiWorkerKerasBenchmarkReal(Resnet50MultiWorkerKerasBenchmark):
  """Resnet50 multi-worker real data benchmark tests."""

  def __init__(self, output_dir=None, root_data_dir=None, **kwargs):
    def_flags = {}
    def_flags['skip_eval'] = True
    def_flags['report_accuracy_metrics'] = False
1150
    def_flags['data_dir'] = os.path.join(root_data_dir, 'imagenet')
1151
1152
1153
1154
1155
1156
1157
    def_flags['train_steps'] = 110
    def_flags['log_steps'] = 10

    super(Resnet50MultiWorkerKerasBenchmarkReal, self).__init__(
        output_dir=output_dir, default_flags=def_flags)


1158
1159
if __name__ == '__main__':
  tf.test.main()