Spaces:
Running
Running
File size: 13,984 Bytes
0b8359d |
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 |
# Copyright 2019 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 BERT benchmarks and accuracy tests."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import functools
import json
import math
import os
import time
# pylint: disable=g-bad-import-order
from absl import flags
from absl.testing import flagsaver
import tensorflow as tf
# pylint: enable=g-bad-import-order
from official.benchmark import bert_benchmark_utils as benchmark_utils
from official.benchmark import owner_utils
from official.nlp.bert import configs
from official.nlp.bert import run_classifier
from official.utils.misc import distribution_utils
from official.benchmark import benchmark_wrappers
# pylint: disable=line-too-long
PRETRAINED_CHECKPOINT_PATH = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16/bert_model.ckpt'
CLASSIFIER_TRAIN_DATA_PATH = 'gs://tf-perfzero-data/bert/classification/mrpc_train.tf_record'
CLASSIFIER_EVAL_DATA_PATH = 'gs://tf-perfzero-data/bert/classification/mrpc_eval.tf_record'
CLASSIFIER_INPUT_META_DATA_PATH = 'gs://tf-perfzero-data/bert/classification/mrpc_meta_data'
MODEL_CONFIG_FILE_PATH = 'gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-24_H-1024_A-16/bert_config.json'
# pylint: enable=line-too-long
TMP_DIR = os.getenv('TMPDIR')
FLAGS = flags.FLAGS
class BertClassifyBenchmarkBase(benchmark_utils.BertBenchmarkBase):
"""Base class to hold methods common to test classes in the module."""
def __init__(self, output_dir=None, tpu=None):
super(BertClassifyBenchmarkBase, self).__init__(output_dir, tpu=tpu)
self.num_epochs = None
self.num_steps_per_epoch = None
FLAGS.steps_per_loop = 1
@flagsaver.flagsaver
def _run_bert_classifier(self, callbacks=None, use_ds=True):
"""Starts BERT classification task."""
with tf.io.gfile.GFile(FLAGS.input_meta_data_path, 'rb') as reader:
input_meta_data = json.loads(reader.read().decode('utf-8'))
bert_config = configs.BertConfig.from_json_file(FLAGS.bert_config_file)
epochs = self.num_epochs if self.num_epochs else FLAGS.num_train_epochs
if self.num_steps_per_epoch:
steps_per_epoch = self.num_steps_per_epoch
else:
train_data_size = input_meta_data['train_data_size']
steps_per_epoch = int(train_data_size / FLAGS.train_batch_size)
warmup_steps = int(epochs * steps_per_epoch * 0.1)
eval_steps = int(
math.ceil(input_meta_data['eval_data_size'] / FLAGS.eval_batch_size))
if self.tpu:
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy='tpu', tpu_address=self.tpu)
else:
strategy = distribution_utils.get_distribution_strategy(
distribution_strategy='mirrored' if use_ds else 'off',
num_gpus=self.num_gpus)
max_seq_length = input_meta_data['max_seq_length']
train_input_fn = run_classifier.get_dataset_fn(
FLAGS.train_data_path,
max_seq_length,
FLAGS.train_batch_size,
is_training=True)
eval_input_fn = run_classifier.get_dataset_fn(
FLAGS.eval_data_path,
max_seq_length,
FLAGS.eval_batch_size,
is_training=False)
_, summary = run_classifier.run_bert_classifier(
strategy,
bert_config,
input_meta_data,
FLAGS.model_dir,
epochs,
steps_per_epoch,
FLAGS.steps_per_loop,
eval_steps,
warmup_steps,
FLAGS.learning_rate,
FLAGS.init_checkpoint,
train_input_fn,
eval_input_fn,
training_callbacks=False,
custom_callbacks=callbacks)
return summary
class BertClassifyBenchmarkReal(BertClassifyBenchmarkBase):
"""Short benchmark performance tests for BERT model.
Tests BERT classification performance in different GPU, TPU configurations.
The naming convention of below test cases follow
`benchmark_(number of gpus)_gpu_(dataset type)` for GPUs and
`benchmark_(topology)_tpu_(dataset type)` for TPUs.
"""
def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
super(BertClassifyBenchmarkReal, self).__init__(
output_dir=output_dir, tpu=tpu)
self.train_data_path = CLASSIFIER_TRAIN_DATA_PATH
self.eval_data_path = CLASSIFIER_EVAL_DATA_PATH
self.bert_config_file = MODEL_CONFIG_FILE_PATH
self.input_meta_data_path = CLASSIFIER_INPUT_META_DATA_PATH
# Since we only care about performance metrics, we limit
# the number of training steps and epochs to prevent unnecessarily
# long tests.
self.num_steps_per_epoch = 100
self.num_epochs = 1
@benchmark_wrappers.enable_runtime_flags
def _run_and_report_benchmark(self,
training_summary_path,
min_accuracy=0,
max_accuracy=1,
use_ds=True):
"""Starts BERT performance benchmark test."""
start_time_sec = time.time()
summary = self._run_bert_classifier(
callbacks=[self.timer_callback], use_ds=use_ds)
wall_time_sec = time.time() - start_time_sec
# Since we do not load from any pretrained checkpoints, we ignore all
# accuracy metrics.
summary.pop('eval_metrics', None)
summary['start_time_sec'] = start_time_sec
super(BertClassifyBenchmarkReal, self)._report_benchmark(
stats=summary,
wall_time_sec=wall_time_sec,
min_accuracy=min_accuracy,
max_accuracy=max_accuracy)
def benchmark_1_gpu_mrpc(self):
"""Test BERT model performance with 1 GPU."""
self._setup()
self.num_gpus = 1
FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_mrpc')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.train_batch_size = 4
FLAGS.eval_batch_size = 4
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path)
def benchmark_1_gpu_mrpc_xla(self):
"""Test BERT model performance with 1 GPU."""
self._setup()
self.num_gpus = 1
FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_mrpc_xla')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.train_batch_size = 4
FLAGS.eval_batch_size = 4
FLAGS.enable_xla = True
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path)
def benchmark_1_gpu_mrpc_no_dist_strat(self):
"""Test BERT model performance with 1 GPU, no distribution strategy."""
self._setup()
self.num_gpus = 1
FLAGS.model_dir = self._get_model_dir('benchmark_1_gpu_mrpc_no_dist_strat')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.train_batch_size = 4
FLAGS.eval_batch_size = 4
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path, use_ds=False)
@owner_utils.Owner('tf-model-garden')
def benchmark_8_gpu_mrpc(self):
"""Test BERT model performance with 8 GPUs."""
self._setup()
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mrpc')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path)
def benchmark_1_gpu_amp_mrpc_no_dist_strat(self):
"""Performance for 1 GPU no DS with automatic mixed precision."""
self._setup()
self.num_gpus = 1
FLAGS.model_dir = self._get_model_dir(
'benchmark_1_gpu_amp_mrpc_no_dist_strat')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.train_batch_size = 4
FLAGS.eval_batch_size = 4
FLAGS.dtype = 'fp16'
FLAGS.fp16_implementation = 'graph_rewrite'
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path, use_ds=False)
def benchmark_8_gpu_amp_mrpc(self):
"""Test BERT model performance with 8 GPUs with automatic mixed precision."""
self._setup()
self.num_gpus = 8
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_amp_mrpc')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.train_batch_size = 32
FLAGS.eval_batch_size = 32
FLAGS.dtype = 'fp16'
FLAGS.fp16_implementation = 'graph_rewrite'
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path, use_ds=False)
@owner_utils.Owner('tf-model-garden')
def benchmark_2x2_tpu_mrpc(self):
"""Test BERT model performance with 2x2 TPU."""
self._setup()
FLAGS.steps_per_loop = 50
FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_mrpc')
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.train_batch_size = 32
FLAGS.eval_batch_size = 32
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path, use_ds=False)
class BertClassifyAccuracy(BertClassifyBenchmarkBase):
"""Short accuracy test for BERT model.
Tests BERT classification task model accuracy. The naming
convention of below test cases follow
`benchmark_(number of gpus)_gpu_(dataset type)` format.
"""
def __init__(self, output_dir=TMP_DIR, tpu=None, **kwargs):
self.train_data_path = CLASSIFIER_TRAIN_DATA_PATH
self.eval_data_path = CLASSIFIER_EVAL_DATA_PATH
self.bert_config_file = MODEL_CONFIG_FILE_PATH
self.input_meta_data_path = CLASSIFIER_INPUT_META_DATA_PATH
self.pretrained_checkpoint_path = PRETRAINED_CHECKPOINT_PATH
super(BertClassifyAccuracy, self).__init__(output_dir=output_dir, tpu=tpu)
@benchmark_wrappers.enable_runtime_flags
def _run_and_report_benchmark(self,
training_summary_path,
min_accuracy=0.84,
max_accuracy=0.88):
"""Starts BERT accuracy benchmark test."""
start_time_sec = time.time()
summary = self._run_bert_classifier(callbacks=[self.timer_callback])
wall_time_sec = time.time() - start_time_sec
super(BertClassifyAccuracy, self)._report_benchmark(
stats=summary,
wall_time_sec=wall_time_sec,
min_accuracy=min_accuracy,
max_accuracy=max_accuracy)
def _setup(self):
super(BertClassifyAccuracy, self)._setup()
FLAGS.train_data_path = self.train_data_path
FLAGS.eval_data_path = self.eval_data_path
FLAGS.input_meta_data_path = self.input_meta_data_path
FLAGS.bert_config_file = self.bert_config_file
FLAGS.init_checkpoint = self.pretrained_checkpoint_path
@owner_utils.Owner('tf-model-garden')
def benchmark_8_gpu_mrpc(self):
"""Run BERT model accuracy test with 8 GPUs.
Due to comparatively small cardinality of MRPC dataset, training
accuracy metric has high variance between trainings. As so, we
set the wide range of allowed accuracy (84% to 88%).
"""
self._setup()
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mrpc')
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path)
def benchmark_8_gpu_mrpc_xla(self):
"""Run BERT model accuracy test with 8 GPUs with XLA."""
self._setup()
FLAGS.model_dir = self._get_model_dir('benchmark_8_gpu_mrpc_xla')
FLAGS.enable_xla = True
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path)
@owner_utils.Owner('tf-model-garden')
def benchmark_2x2_tpu_mrpc(self):
"""Run BERT model accuracy test on 2x2 TPU."""
self._setup()
FLAGS.steps_per_loop = 50
FLAGS.model_dir = self._get_model_dir('benchmark_2x2_tpu_mrpc')
summary_path = os.path.join(FLAGS.model_dir,
'summaries/training_summary.txt')
self._run_and_report_benchmark(summary_path)
if __name__ == '__main__':
tf.test.main()
|