File size: 16,269 Bytes
97b6013 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 |
# 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
from absl import logging
from absl.testing import flagsaver
import tensorflow as tf
from official.benchmark import benchmark_wrappers
from official.benchmark import owner_utils
from official.benchmark.perfzero_benchmark import PerfZeroBenchmark
from official.recommendation import ncf_common
from official.recommendation import ncf_keras_main
from official.utils.flags import core
FLAGS = flags.FLAGS
NCF_DATA_DIR_NAME = 'movielens_data'
NCF_TF_REGRESSION_DATA_DIR_NAME = 'gs://tf-regression/ncf/data'
class NCFKerasBenchmarkBase(PerfZeroBenchmark):
"""Base class for NCF model benchmark."""
def __init__(self, output_dir=None, default_flags=None, **kwargs):
super(NCFKerasBenchmarkBase, self).__init__(output_dir, default_flags,
**kwargs)
# Run all benchmarks with ml_perf flag.
self.default_flags['ml_perf'] = True
def _setup(self):
"""Sets up and resets flags before each test."""
logging.set_verbosity(logging.INFO)
if NCFKerasBenchmarkBase.local_flags is None:
ncf_common.define_ncf_flags()
# 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()
NCFKerasBenchmarkBase.local_flags = saved_flag_values
else:
flagsaver.restore_flag_values(NCFKerasBenchmarkBase.local_flags)
@benchmark_wrappers.enable_runtime_flags
def _run_and_report_benchmark(self, hr_at_10_min=0, hr_at_10_max=0):
start_time_sec = time.time()
stats = ncf_keras_main.run_ncf(FLAGS)
wall_time_sec = time.time() - start_time_sec
metrics = []
metrics.append({'name': 'exp_per_second',
'value': stats['avg_exp_per_second']})
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)
class NCFKerasAccuracy(NCFKerasBenchmarkBase):
"""Benchmark NCF model using real data."""
def __init__(self,
output_dir=None,
root_data_dir=None,
default_flags=None,
**kwargs):
root_data_dir = root_data_dir if root_data_dir else ''
default_flags = {}
default_flags['dataset'] = 'ml-20m'
default_flags['num_gpus'] = 1
default_flags['train_epochs'] = 10
default_flags['clean'] = True
default_flags['batch_size'] = 99000
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['ml_perf'] = True
default_flags['use_synthetic_data'] = False
default_flags['data_dir'] = os.path.join(root_data_dir, NCF_DATA_DIR_NAME)
super(NCFKerasAccuracy, self).__init__(
output_dir=output_dir,
default_flags=default_flags,
**kwargs)
def _run_and_report_benchmark_mlperf_like(self):
"""Run test and report results.
Note: MLPerf like tests are not tuned to hit a specific hr@10 value, but
we want it recorded.
"""
self._run_and_report_benchmark(hr_at_10_min=0.61)
def _run_and_report_benchmark(self, hr_at_10_min=0.630, hr_at_10_max=0.645):
"""Run test and report results.
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(
hr_at_10_min=hr_at_10_min,
hr_at_10_max=hr_at_10_max)
def _set_8_gpu_defaults(self):
FLAGS.num_gpus = 8
FLAGS.learning_rate = 0.0045
FLAGS.beta1 = 0.25
FLAGS.beta2 = 0.5
FLAGS.epsilon = 1e-8
FLAGS.train_epochs = 14
FLAGS.batch_size = 99000
FLAGS.eval_batch_size = 160000
FLAGS.train_dataset_path = os.path.join(NCF_TF_REGRESSION_DATA_DIR_NAME,
'training_cycle_*/*')
FLAGS.eval_dataset_path = os.path.join(NCF_TF_REGRESSION_DATA_DIR_NAME,
'eval_data/*')
FLAGS.input_meta_data_path = os.path.join(NCF_TF_REGRESSION_DATA_DIR_NAME,
'metadata')
FLAGS.data_dir = NCF_TF_REGRESSION_DATA_DIR_NAME
def benchmark_1_gpu_early_stop(self):
self._setup()
FLAGS.early_stopping = True
self._run_and_report_benchmark()
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()
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()
def benchmark_1_gpu_ctl_early_stop(self):
self._setup()
FLAGS.keras_use_ctl = True
FLAGS.early_stopping = True
self._run_and_report_benchmark()
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()
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()
def benchmark_2_gpus_early_stop(self):
self._setup()
FLAGS.early_stopping = True
FLAGS.num_gpus = 2
FLAGS.eval_batch_size = 160000
self._run_and_report_benchmark()
def benchmark_2_gpus_ctl_early_stop(self):
"""NCF with custom training loop. Works only in TF 2.0."""
self._setup()
FLAGS.keras_use_ctl = True
FLAGS.early_stopping = True
FLAGS.num_gpus = 2
FLAGS.eval_batch_size = 160000
self._run_and_report_benchmark()
#############################################
# Tests below with mlperf in the test name are of two types:
# 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.
#############################################
def benchmark_1_gpu_mlperf_like(self):
"""1 GPU using keras fit/compile."""
self._setup()
FLAGS.train_epochs = 7
self._run_and_report_benchmark_mlperf_like()
def benchmark_1_gpu_no_dist_strat_mlperf_like(self):
"""1 GPU using compile/fit without dist_strat."""
self._setup()
FLAGS.train_epochs = 7
FLAGS.distribution_strategy = 'off'
self._run_and_report_benchmark_mlperf_like()
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
self._run_and_report_benchmark_mlperf_like()
def benchmark_xla_1_gpu_mlperf_like(self):
"""1 GPU using compile/fit with XLA."""
self._setup()
FLAGS.train_epochs = 7
FLAGS.enable_xla = True
self._run_and_report_benchmark_mlperf_like()
def benchmark_1_gpu_ctl_mlperf_like(self):
"""1 GPU using CTL."""
self._setup()
FLAGS.keras_use_ctl = True
FLAGS.train_epochs = 7
self._run_and_report_benchmark_mlperf_like()
def benchmark_1_gpu_ctl_fp16_mlperf_like(self):
"""1 GPU using CTL and FP16."""
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()
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()
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()
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()
def benchmark_xla_1_gpu_ctl_mlperf_like(self):
"""1 GPU using CTL with XLA."""
self._setup()
FLAGS.keras_use_ctl = True
FLAGS.enable_xla = True
FLAGS.train_epochs = 7
self._run_and_report_benchmark_mlperf_like()
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()
def benchmark_xla_1_gpu_ctl_fp16_mlperf_like(self):
"""1 GPU using CTL with XLA and FP16."""
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()
def benchmark_8_gpu_mlperf_like(self):
"""8 GPU using keras fit/compile."""
self._setup()
FLAGS.num_gpus = 8
FLAGS.train_epochs = 17
FLAGS.batch_size = 1048576
FLAGS.eval_batch_size = 160000
FLAGS.learning_rate = 0.0045
FLAGS.beta1 = 0.25
FLAGS.beta2 = 0.5
FLAGS.epsilon = 1e-8
self._run_and_report_benchmark_mlperf_like()
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
FLAGS.eval_batch_size = 160000
FLAGS.learning_rate = 0.0045
FLAGS.beta1 = 0.25
FLAGS.beta2 = 0.5
FLAGS.epsilon = 1e-8
self._run_and_report_benchmark_mlperf_like()
def benchmark_8_gpu_tf_data_ctl_mlperf_like(self):
"""8 GPU using CTL."""
self._setup()
self._set_8_gpu_defaults()
FLAGS.keras_use_ctl = True
self._run_and_report_benchmark_mlperf_like()
def benchmark_8_gpu_tf_data_fp16_mlperf_like(self):
"""8 GPU FP16."""
self._setup()
self._set_8_gpu_defaults()
FLAGS.dtype = 'fp16'
FLAGS.loss_scale = 8192
self._run_and_report_benchmark_mlperf_like()
def benchmark_8_gpu_tf_data_ctl_fp16_mlperf_like(self):
"""8 GPU FP16 using CTL."""
self._setup()
self._set_8_gpu_defaults()
FLAGS.keras_use_ctl = True
FLAGS.dtype = 'fp16'
FLAGS.loss_scale = 8192
self._run_and_report_benchmark_mlperf_like()
def benchmark_8_gpu_tf_data_ctl_fp16_graph_rewrite_mlperf_like(self):
"""8 GPU FP16 graph rewrite using CTL."""
self._setup()
self._set_8_gpu_defaults()
FLAGS.keras_use_ctl = True
FLAGS.dtype = 'fp16'
FLAGS.fp16_implementation = 'graph_rewrite'
FLAGS.loss_scale = 8192
self._run_and_report_benchmark_mlperf_like()
class NCFKerasBenchmarkReal(NCFKerasBenchmarkBase):
"""NCF Keras throughput benchmarks."""
def __init__(self,
output_dir=None,
root_data_dir=None,
default_flags=None,
**kwargs):
root_data_dir = root_data_dir if root_data_dir else ''
default_flags = {}
default_flags['dataset'] = 'ml-20m'
default_flags['num_gpus'] = 1
default_flags['train_epochs'] = 14
default_flags['clean'] = True
default_flags['batch_size'] = 99000
default_flags['eval_batch_size'] = 160000
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['ml_perf'] = True
default_flags['use_synthetic_data'] = False
default_flags['train_dataset_path'] = os.path.join(
NCF_TF_REGRESSION_DATA_DIR_NAME, 'training_cycle_*/*')
default_flags['eval_dataset_path'] = os.path.join(
NCF_TF_REGRESSION_DATA_DIR_NAME, 'eval_data/*')
default_flags['input_meta_data_path'] = os.path.join(
NCF_TF_REGRESSION_DATA_DIR_NAME, 'metadata')
default_flags['data_dir'] = NCF_TF_REGRESSION_DATA_DIR_NAME
super(NCFKerasBenchmarkReal, self).__init__(
output_dir=output_dir, default_flags=default_flags, **kwargs)
def benchmark_2x2_tpu(self):
"""2x2 TPU using CTL with distribution strategy."""
self._setup()
FLAGS.distribution_strategy = 'tpu'
FLAGS.keras_use_ctl = True
FLAGS.num_gpus = 0
FLAGS.train_epochs = 1
self._run_and_report_benchmark()
@owner_utils.Owner('tf-graph-compiler')
def benchmark_2x2_tpu_mlir(self):
"""2x2 TPU using CTL with distribution strategy using the MLIR bridge."""
self._setup()
FLAGS.distribution_strategy = 'tpu'
FLAGS.keras_use_ctl = True
FLAGS.num_gpus = 0
FLAGS.train_epochs = 1
tf.config.experimental.enable_mlir_bridge()
self._run_and_report_benchmark()
class NCFKerasSynth(NCFKerasBenchmarkBase):
"""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
default_flags['train_epochs'] = 8
default_flags['batch_size'] = 99000
default_flags['eval_batch_size'] = 160000
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
super(NCFKerasSynth, self).__init__(
output_dir=output_dir,
default_flags=default_flags,
**kwargs)
def benchmark_1_gpu(self):
self._setup()
self._run_and_report_benchmark()
def benchmark_2_gpus(self):
self._setup()
FLAGS.num_gpus = 2
self._run_and_report_benchmark()
if __name__ == '__main__':
tf.test.main()
|