ncf_keras_benchmark.py 15 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# 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 absolute_import
from __future__ import division
from __future__ import print_function

import os
import time

from absl import flags
24
from absl import logging
25
from absl.testing import flagsaver
Hongkun Yu's avatar
Hongkun Yu committed
26
import tensorflow as tf
27
from official.benchmark import benchmark_wrappers
28
29
30
31
32
from official.recommendation import ncf_common
from official.recommendation import ncf_keras_main
from official.utils.flags import core

FLAGS = flags.FLAGS
Toby Boyd's avatar
Toby Boyd committed
33
NCF_DATA_DIR_NAME = 'movielens_data'
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
34
NCF_TF_DATA_1M_BATCH_DIR_NAME = 'gs://tf-perfzero-data/movielens_data/ncf_8gpu_1M_batch'
Toby Boyd's avatar
Toby Boyd committed
35

36

37
class NCFKerasBenchmarkBase(tf.test.Benchmark):
38
39
40
41
42
43
44
45
46
  """Base class for NCF model benchmark."""
  local_flags = None

  def __init__(self,
               output_dir=None,
               default_flags=None,
               **kwargs):
    self.output_dir = output_dir
    self.default_flags = default_flags or {}
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
47
48
    # Run all benchmarks with ml_perf flag.
    self.default_flags['ml_perf'] = True
49
50
51

  def _setup(self):
    """Sets up and resets flags before each test."""
52
    logging.set_verbosity(logging.INFO)
53
    if NCFKerasBenchmarkBase.local_flags is None:
Toby Boyd's avatar
Toby Boyd committed
54
      ncf_common.define_ncf_flags()
55
56
57
58
      # Loads flags to get defaults to then override. List cannot be empty.
      flags.FLAGS(['foo'])
      core.set_defaults(**self.default_flags)
      saved_flag_values = flagsaver.save_flag_values()
59
      NCFKerasBenchmarkBase.local_flags = saved_flag_values
60
    else:
61
      flagsaver.restore_flag_values(NCFKerasBenchmarkBase.local_flags)
62

63
  @benchmark_wrappers.enable_runtime_flags
Toby Boyd's avatar
Toby Boyd committed
64
  def _run_and_report_benchmark(self, hr_at_10_min=0, hr_at_10_max=0):
65
66
67
68
    start_time_sec = time.time()
    stats = ncf_keras_main.run_ncf(FLAGS)
    wall_time_sec = time.time() - start_time_sec

Toby Boyd's avatar
Toby Boyd committed
69
70
71
    metrics = []
    metrics.append({'name': 'exp_per_second',
                    'value': stats['avg_exp_per_second']})
72

Toby Boyd's avatar
Toby Boyd committed
73
74
75
76
77
78
79
80
81
82
    if hr_at_10_min > 0:
      metrics.append({'name': 'hr_at_10',
                      'value': stats['eval_hit_rate'],
                      'min_value': hr_at_10_min,
                      'max_value': hr_at_10_max})

      metrics.append({'name': 'train_loss',
                      'value': stats['loss']})

    self.report_benchmark(iters=-1, wall_time=wall_time_sec, metrics=metrics)
83
84


85
class NCFKerasAccuracy(NCFKerasBenchmarkBase):
86
87
88
89
  """Benchmark NCF model using real data."""

  def __init__(self,
               output_dir=None,
Toby Boyd's avatar
Toby Boyd committed
90
               root_data_dir=None,
91
92
               default_flags=None,
               **kwargs):
Hongkun Yu's avatar
Hongkun Yu committed
93
    root_data_dir = root_data_dir if root_data_dir else ''
94
95
96
    default_flags = {}
    default_flags['dataset'] = 'ml-20m'
    default_flags['num_gpus'] = 1
97
    default_flags['train_epochs'] = 10
98
    default_flags['clean'] = True
99
    default_flags['batch_size'] = 99000
100
101
102
103
104
105
106
    default_flags['learning_rate'] = 0.00382059
    default_flags['beta1'] = 0.783529
    default_flags['beta2'] = 0.909003
    default_flags['epsilon'] = 1.45439e-07
    default_flags['layers'] = [256, 256, 128, 64]
    default_flags['num_factors'] = 64
    default_flags['hr_threshold'] = 0.635
107
    default_flags['ml_perf'] = True
108
    default_flags['use_synthetic_data'] = False
Toby Boyd's avatar
Toby Boyd committed
109
    default_flags['data_dir'] = os.path.join(root_data_dir, NCF_DATA_DIR_NAME)
110

111
    super(NCFKerasAccuracy, self).__init__(
112
113
114
115
        output_dir=output_dir,
        default_flags=default_flags,
        **kwargs)

Toby Boyd's avatar
Toby Boyd committed
116
117
  def _run_and_report_benchmark_mlperf_like(self):
    """Run test and report results.
Toby Boyd's avatar
Toby Boyd committed
118

Toby Boyd's avatar
Toby Boyd committed
119
120
121
    Note: MLPerf like tests are not tuned to hit a specific hr@10 value, but
    we want it recorded.
    """
122
    self._run_and_report_benchmark(hr_at_10_min=0.61)
Toby Boyd's avatar
Toby Boyd committed
123

124
  def _run_and_report_benchmark(self, hr_at_10_min=0.630, hr_at_10_max=0.645):
Toby Boyd's avatar
Toby Boyd committed
125
    """Run test and report results.
Toby Boyd's avatar
Toby Boyd committed
126

Toby Boyd's avatar
Toby Boyd committed
127
128
129
130
131
132
133
134
    Note: Target is 0.635, but some runs are below that level. Until we have
    multi-run tests, we have to accept a lower target.

    Args:
      hr_at_10_min: Minimum acceptable hr@10 value.
      hr_at_10_max: Maximum acceptable hr@10 value.
    """
    super(NCFKerasAccuracy, self)._run_and_report_benchmark(
135
136
        hr_at_10_min=hr_at_10_min,
        hr_at_10_max=hr_at_10_max)
137

138
  def benchmark_1_gpu_early_stop(self):
139
    self._setup()
140
    FLAGS.early_stopping = True
141
142
    self._run_and_report_benchmark()

143
144
145
146
147
148
  def benchmark_1_gpu_no_dist_strat_early_stop(self):
    self._setup()
    FLAGS.distribution_strategy = 'off'
    FLAGS.early_stopping = True
    self._run_and_report_benchmark()

149
150
151
152
153
154
155
156
157
158
159
160
161
  def benchmark_1_gpu_no_dist_strat_run_eagerly_early_stop(self):
    self._setup()
    FLAGS.distribution_strategy = 'off'
    FLAGS.early_stopping = True
    FLAGS.run_eagerly = True
    self._run_and_report_benchmark()

  def benchmark_xla_1_gpu_early_stop(self):
    self._setup()
    FLAGS.early_stopping = True
    FLAGS.enable_xla = True
    self._run_and_report_benchmark()

162
163
164
165
166
167
  def benchmark_1_gpu_ctl_early_stop(self):
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.early_stopping = True
    self._run_and_report_benchmark()

168
169
170
171
172
173
174
  def benchmark_1_gpu_ctl_run_eagerly_early_stop(self):
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.early_stopping = True
    FLAGS.run_eagerly = True
    self._run_and_report_benchmark()

175
176
177
178
179
180
181
  def benchmark_xla_1_gpu_ctl_early_stop(self):
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.early_stopping = True
    FLAGS.enable_xla = True
    self._run_and_report_benchmark()

182
183
184
185
  def benchmark_2_gpus_early_stop(self):
    self._setup()
    FLAGS.early_stopping = True
    FLAGS.num_gpus = 2
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
186
    FLAGS.eval_batch_size = 160000
187
    self._run_and_report_benchmark()
188

189
  def benchmark_2_gpus_ctl_early_stop(self):
190
    """NCF with custom training loop. Works only in TF 2.0."""
191
192
193
194
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.early_stopping = True
    FLAGS.num_gpus = 2
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
195
    FLAGS.eval_batch_size = 160000
196
197
    self._run_and_report_benchmark()

198
#############################################
199
# Tests below with mlperf in the test name are of two types:
200
201
202
203
204
205
206
#  1) 1 GPU tests are based on MLPerf 0.5 and the TensorFlow pulled submission.
#  2) 8 GPU tests are based on MLPerf 0.5 and use NVIDIA's hyper parameters.
#
# The purpose of both is to get a number to compare to existing results. To do
# this the number of epochs is held constant rather than a race to a given
# accuracy. The accuracy validation is done by the "early_stop" tests.
#############################################
207
208

  def benchmark_1_gpu_mlperf_like(self):
209
    """1 GPU using keras fit/compile."""
210
211
    self._setup()
    FLAGS.train_epochs = 7
Toby Boyd's avatar
Toby Boyd committed
212
    self._run_and_report_benchmark_mlperf_like()
213
214

  def benchmark_1_gpu_no_dist_strat_mlperf_like(self):
215
    """1 GPU using compile/fit without dist_strat."""
216
217
218
    self._setup()
    FLAGS.train_epochs = 7
    FLAGS.distribution_strategy = 'off'
Toby Boyd's avatar
Toby Boyd committed
219
    self._run_and_report_benchmark_mlperf_like()
220
221
222
223
224
225

  def benchmark_1_gpu_no_dist_strat_run_eagerly_mlperf_like(self):
    self._setup()
    FLAGS.train_epochs = 7
    FLAGS.distribution_strategy = 'off'
    FLAGS.run_eagerly = True
Toby Boyd's avatar
Toby Boyd committed
226
    self._run_and_report_benchmark_mlperf_like()
227
228

  def benchmark_xla_1_gpu_mlperf_like(self):
229
    """1 GPU using compile/fit with XLA."""
230
231
    self._setup()
    FLAGS.train_epochs = 7
232
    FLAGS.enable_xla = True
Toby Boyd's avatar
Toby Boyd committed
233
    self._run_and_report_benchmark_mlperf_like()
234

235
236
237
238
239
  def benchmark_1_gpu_ctl_mlperf_like(self):
    """1 GPU using CTL."""
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.train_epochs = 7
Toby Boyd's avatar
Toby Boyd committed
240
    self._run_and_report_benchmark_mlperf_like()
241

Nimit Nigania's avatar
Nimit Nigania committed
242
  def benchmark_1_gpu_ctl_fp16_mlperf_like(self):
Tomasz Grel's avatar
Tomasz Grel committed
243
    """1 GPU using CTL and FP16."""
Nimit Nigania's avatar
Nimit Nigania committed
244
245
246
247
248
249
250
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.train_epochs = 7
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 8192
    self._run_and_report_benchmark_mlperf_like()

Tomasz Grel's avatar
Tomasz Grel committed
251
252
253
254
255
256
257
258
  def benchmark_1_gpu_fp16_mlperf_like(self):
    """1 GPU using FP16."""
    self._setup()
    FLAGS.train_epochs = 7
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 8192
    self._run_and_report_benchmark_mlperf_like()

259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
  def benchmark_1_gpu_ctl_fp16_graph_rewrite_mlperf_like(self):
    """1 GPU using CTL and FP16 graph rewrite."""
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.train_epochs = 7
    FLAGS.dtype = 'fp16'
    FLAGS.fp16_implementation = 'graph_rewrite'
    FLAGS.loss_scale = 8192
    self._run_and_report_benchmark_mlperf_like()

  def benchmark_1_gpu_fp16_graph_rewrite_mlperf_like(self):
    """1 GPU using FP16 graph rewrite."""
    self._setup()
    FLAGS.train_epochs = 7
    FLAGS.dtype = 'fp16'
    FLAGS.fp16_implementation = 'graph_rewrite'
    FLAGS.loss_scale = 8192
    self._run_and_report_benchmark_mlperf_like()

278
279
280
281
282
283
284
285
  def benchmark_1_gpu_ctl_run_eagerly_mlperf_like(self):
    """1 GPU using CTL with eager and distribution strategy."""
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.run_eagerly = True
    FLAGS.train_epochs = 7
    self._run_and_report_benchmark()

286
287
  def benchmark_xla_1_gpu_ctl_mlperf_like(self):
    """1 GPU using CTL with XLA."""
288
289
    self._setup()
    FLAGS.keras_use_ctl = True
290
291
    FLAGS.enable_xla = True
    FLAGS.train_epochs = 7
Toby Boyd's avatar
Toby Boyd committed
292
    self._run_and_report_benchmark_mlperf_like()
293

Tomasz Grel's avatar
Tomasz Grel committed
294
295
296
297
298
299
300
301
302
  def benchmark_xla_1_gpu_fp16_mlperf_like(self):
    """1 GPU using with XLA and FP16."""
    self._setup()
    FLAGS.enable_xla = True
    FLAGS.train_epochs = 7
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 8192
    self._run_and_report_benchmark_mlperf_like()

Nimit Nigania's avatar
Nimit Nigania committed
303
  def benchmark_xla_1_gpu_ctl_fp16_mlperf_like(self):
Tomasz Grel's avatar
Tomasz Grel committed
304
    """1 GPU using CTL with XLA and FP16."""
Nimit Nigania's avatar
Nimit Nigania committed
305
306
307
308
309
310
311
312
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.enable_xla = True
    FLAGS.train_epochs = 7
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 8192
    self._run_and_report_benchmark_mlperf_like()

313
314
315
  def benchmark_8_gpu_mlperf_like(self):
    """8 GPU using keras fit/compile."""
    self._setup()
316
317
318
    FLAGS.num_gpus = 8
    FLAGS.train_epochs = 17
    FLAGS.batch_size = 1048576
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
319
    FLAGS.eval_batch_size = 160000
320
321
322
323
    FLAGS.learning_rate = 0.0045
    FLAGS.beta1 = 0.25
    FLAGS.beta2 = 0.5
    FLAGS.epsilon = 1e-8
Toby Boyd's avatar
Toby Boyd committed
324
    self._run_and_report_benchmark_mlperf_like()
325

326
327
328
329
330
331
332
  def benchmark_8_gpu_ctl_mlperf_like(self):
    """8 GPU using CTL."""
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.num_gpus = 8
    FLAGS.train_epochs = 17
    FLAGS.batch_size = 1048576
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
333
    FLAGS.eval_batch_size = 160000
334
335
336
337
    FLAGS.learning_rate = 0.0045
    FLAGS.beta1 = 0.25
    FLAGS.beta2 = 0.5
    FLAGS.epsilon = 1e-8
Toby Boyd's avatar
Toby Boyd committed
338
    self._run_and_report_benchmark_mlperf_like()
339

A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
  def benchmark_8_gpu_tf_data_ctl_mlperf_like(self):
    """8 GPU using CTL."""
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.num_gpus = 8
    FLAGS.train_epochs = 17
    FLAGS.batch_size = 1048576
    FLAGS.eval_batch_size = 1048000
    FLAGS.learning_rate = 0.0045
    FLAGS.beta1 = 0.25
    FLAGS.beta2 = 0.5
    FLAGS.epsilon = 1e-8
    FLAGS.train_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "training_cycle_*/*")
    FLAGS.eval_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "eval_data/*")
    FLAGS.input_meta_data_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "meta_data.json")
    self._run_and_report_benchmark_mlperf_like()

Tomasz Grel's avatar
Tomasz Grel committed
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
  def benchmark_8_gpu_tf_data_fp16_mlperf_like(self):
    """8 GPU FP16"""
    self._setup()
    FLAGS.num_gpus = 8
    FLAGS.train_epochs = 17
    FLAGS.batch_size = 1048576
    FLAGS.eval_batch_size = 1048000
    FLAGS.learning_rate = 0.0045
    FLAGS.beta1 = 0.25
    FLAGS.beta2 = 0.5
    FLAGS.epsilon = 1e-8
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 8192
    FLAGS.train_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "training_cycle_*/*")
    FLAGS.eval_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "eval_data/*")
    FLAGS.input_meta_data_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "meta_data.json")
    self._run_and_report_benchmark_mlperf_like()

A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
375
  def benchmark_8_gpu_tf_data_ctl_fp16_mlperf_like(self):
Tomasz Grel's avatar
Tomasz Grel committed
376
    """8 GPU FP16 using CTL"""
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.num_gpus = 8
    FLAGS.train_epochs = 17
    FLAGS.batch_size = 1048576
    FLAGS.eval_batch_size = 1048000
    FLAGS.learning_rate = 0.0045
    FLAGS.beta1 = 0.25
    FLAGS.beta2 = 0.5
    FLAGS.epsilon = 1e-8
    FLAGS.dtype = 'fp16'
    FLAGS.loss_scale = 8192
    FLAGS.train_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "training_cycle_*/*")
    FLAGS.eval_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "eval_data/*")
    FLAGS.input_meta_data_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME, "meta_data.json")
    self._run_and_report_benchmark_mlperf_like()
393

394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
  def benchmark_8_gpu_tf_data_ctl_fp16_graph_rewrite_mlperf_like(self):
    """8 GPU FP16 graph rewrite using CTL."""
    self._setup()
    FLAGS.keras_use_ctl = True
    FLAGS.num_gpus = 8
    FLAGS.train_epochs = 17
    FLAGS.batch_size = 1048576
    FLAGS.eval_batch_size = 1048000
    FLAGS.learning_rate = 0.0045
    FLAGS.beta1 = 0.25
    FLAGS.beta2 = 0.5
    FLAGS.epsilon = 1e-8
    FLAGS.dtype = 'fp16'
    FLAGS.fp16_implementation = 'graph_rewrite'
    FLAGS.loss_scale = 8192
    FLAGS.train_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME,
                                            'training_cycle_*/*')
    FLAGS.eval_dataset_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME,
                                           'eval_data/*')
    FLAGS.input_meta_data_path = os.path.join(NCF_TF_DATA_1M_BATCH_DIR_NAME,
                                              'meta_data.json')
    self._run_and_report_benchmark_mlperf_like()


418
class NCFKerasSynth(NCFKerasBenchmarkBase):
419
420
421
422
423
424
425
426
427
428
  """Benchmark NCF model using synthetic data."""

  def __init__(self,
               output_dir=None,
               default_flags=None,
               **kwargs):

    default_flags = {}
    default_flags['dataset'] = 'ml-20m'
    default_flags['num_gpus'] = 1
429
430
    default_flags['train_epochs'] = 8
    default_flags['batch_size'] = 99000
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
431
    default_flags['eval_batch_size'] = 160000
432
433
434
435
436
437
438
439
440
    default_flags['learning_rate'] = 0.00382059
    default_flags['beta1'] = 0.783529
    default_flags['beta2'] = 0.909003
    default_flags['epsilon'] = 1.45439e-07
    default_flags['layers'] = [256, 256, 128, 64]
    default_flags['num_factors'] = 64
    default_flags['hr_threshold'] = 0.635
    default_flags['use_synthetic_data'] = True

441
    super(NCFKerasSynth, self).__init__(
442
443
444
445
446
447
448
        output_dir=output_dir,
        default_flags=default_flags,
        **kwargs)

  def benchmark_1_gpu(self):
    self._setup()
    self._run_and_report_benchmark()
449
450
451
452
453

  def benchmark_2_gpus(self):
    self._setup()
    FLAGS.num_gpus = 2
    self._run_and_report_benchmark()
David Chen's avatar
David Chen committed
454
455
456
457


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