forked from tensorflow/models
-
Notifications
You must be signed in to change notification settings - Fork 0
/
common.py
407 lines (347 loc) · 17.3 KB
/
common.py
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
# 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.
# ==============================================================================
"""Common util functions and classes used by both keras cifar and imagenet."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import multiprocessing
import os
from absl import flags
import numpy as np
import tensorflow as tf
from tensorflow.python.keras.optimizer_v2 import gradient_descent as gradient_descent_v2
from official.utils.flags import core as flags_core
from official.utils.misc import keras_utils
FLAGS = flags.FLAGS
BASE_LEARNING_RATE = 0.1 # This matches Jing's version.
TRAIN_TOP_1 = 'training_accuracy_top_1'
class LearningRateBatchScheduler(tf.keras.callbacks.Callback):
"""Callback to update learning rate on every batch (not epoch boundaries).
N.B. Only support Keras optimizers, not TF optimizers.
Args:
schedule: a function that takes an epoch index and a batch index as input
(both integer, indexed from 0) and returns a new learning rate as
output (float).
"""
def __init__(self, schedule, batch_size, num_images):
super(LearningRateBatchScheduler, self).__init__()
self.schedule = schedule
self.batches_per_epoch = num_images / batch_size
self.batch_size = batch_size
self.epochs = -1
self.prev_lr = -1
def on_epoch_begin(self, epoch, logs=None):
if not hasattr(self.model.optimizer, 'learning_rate'):
raise ValueError('Optimizer must have a "learning_rate" attribute.')
self.epochs += 1
def on_batch_begin(self, batch, logs=None):
"""Executes before step begins."""
lr = self.schedule(self.epochs,
batch,
self.batches_per_epoch,
self.batch_size)
if not isinstance(lr, (float, np.float32, np.float64)):
raise ValueError('The output of the "schedule" function should be float.')
if lr != self.prev_lr:
self.model.optimizer.learning_rate = lr # lr should be a float here
self.prev_lr = lr
tf.compat.v1.logging.debug(
'Epoch %05d Batch %05d: LearningRateBatchScheduler '
'change learning rate to %s.', self.epochs, batch, lr)
class PiecewiseConstantDecayWithWarmup(
tf.keras.optimizers.schedules.LearningRateSchedule):
"""Piecewise constant decay with warmup schedule."""
def __init__(self, batch_size, epoch_size, warmup_epochs, boundaries,
multipliers, compute_lr_on_cpu=True, name=None):
super(PiecewiseConstantDecayWithWarmup, self).__init__()
if len(boundaries) != len(multipliers) - 1:
raise ValueError('The length of boundaries must be 1 less than the '
'length of multipliers')
base_lr_batch_size = 256
num_batches_per_epoch = epoch_size // batch_size
self.rescaled_lr = BASE_LEARNING_RATE * batch_size / base_lr_batch_size
self.step_boundaries = [float(num_batches_per_epoch) * x
for x in boundaries]
self.lr_values = [self.rescaled_lr * m for m in multipliers]
self.warmup_steps = warmup_epochs * num_batches_per_epoch
self.compute_lr_on_cpu = compute_lr_on_cpu
self.name = name
self.learning_rate_ops_cache = {}
def __call__(self, step):
if tf.executing_eagerly():
return self._get_learning_rate(step)
# In an eager function or graph, the current implementation of optimizer
# repeatedly call and thus create ops for the learning rate schedule. To
# avoid this, we cache the ops if not executing eagerly.
graph = tf.compat.v1.get_default_graph()
if graph not in self.learning_rate_ops_cache:
if self.compute_lr_on_cpu:
with tf.device('/device:CPU:0'):
self.learning_rate_ops_cache[graph] = self._get_learning_rate(step)
else:
self.learning_rate_ops_cache[graph] = self._get_learning_rate(step)
return self.learning_rate_ops_cache[graph]
def _get_learning_rate(self, step):
"""Compute learning rate at given step."""
with tf.compat.v1.name_scope(self.name, 'PiecewiseConstantDecayWithWarmup',
[self.rescaled_lr, self.step_boundaries,
self.lr_values, self.warmup_steps,
self.compute_lr_on_cpu]):
def warmup_lr(step):
return self.rescaled_lr * (
tf.cast(step, tf.float32) / tf.cast(self.warmup_steps, tf.float32))
def piecewise_lr(step):
return tf.compat.v1.train.piecewise_constant(
step, self.step_boundaries, self.lr_values)
return tf.cond(step < self.warmup_steps,
lambda: warmup_lr(step),
lambda: piecewise_lr(step))
def get_config(self):
return {
'rescaled_lr': self.rescaled_lr,
'step_boundaries': self.step_boundaries,
'lr_values': self.lr_values,
'warmup_steps': self.warmup_steps,
'compute_lr_on_cpu': self.compute_lr_on_cpu,
'name': self.name
}
def set_gpu_thread_mode_and_count(flags_obj):
"""Set GPU thread mode and count, and adjust dataset threads count."""
cpu_count = multiprocessing.cpu_count()
tf.compat.v1.logging.info('Logical CPU cores: %s', cpu_count)
# Allocate private thread pool for each GPU to schedule and launch kernels
per_gpu_thread_count = flags_obj.per_gpu_thread_count or 2
os.environ['TF_GPU_THREAD_MODE'] = flags_obj.tf_gpu_thread_mode
os.environ['TF_GPU_THREAD_COUNT'] = str(per_gpu_thread_count)
tf.compat.v1.logging.info('TF_GPU_THREAD_COUNT: %s',
os.environ['TF_GPU_THREAD_COUNT'])
tf.compat.v1.logging.info('TF_GPU_THREAD_MODE: %s',
os.environ['TF_GPU_THREAD_MODE'])
# Limit data preprocessing threadpool to CPU cores minus number of total GPU
# private threads and memory copy threads.
total_gpu_thread_count = per_gpu_thread_count * flags_obj.num_gpus
num_runtime_threads = flags_obj.num_gpus
if not flags_obj.datasets_num_private_threads:
flags_obj.datasets_num_private_threads = min(
cpu_count - total_gpu_thread_count - num_runtime_threads,
flags_obj.num_gpus * 8)
tf.compat.v1.logging.info('Set datasets_num_private_threads to %s',
flags_obj.datasets_num_private_threads)
def get_optimizer(learning_rate=0.1):
"""Returns optimizer to use."""
# The learning_rate is overwritten at the beginning of each step by callback.
return gradient_descent_v2.SGD(learning_rate=learning_rate, momentum=0.9)
def get_callbacks(learning_rate_schedule_fn, num_images):
"""Returns common callbacks."""
time_callback = keras_utils.TimeHistory(FLAGS.batch_size, FLAGS.log_steps)
callbacks = [time_callback]
if not FLAGS.use_tensor_lr:
lr_callback = LearningRateBatchScheduler(
learning_rate_schedule_fn,
batch_size=FLAGS.batch_size,
num_images=num_images)
callbacks.append(lr_callback)
if FLAGS.enable_tensorboard:
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=FLAGS.model_dir)
callbacks.append(tensorboard_callback)
if FLAGS.profile_steps:
profiler_callback = keras_utils.get_profiler_callback(
FLAGS.model_dir,
FLAGS.profile_steps,
FLAGS.enable_tensorboard)
callbacks.append(profiler_callback)
return callbacks
def build_stats(history, eval_output, callbacks):
"""Normalizes and returns dictionary of stats.
Args:
history: Results of the training step. Supports both categorical_accuracy
and sparse_categorical_accuracy.
eval_output: Output of the eval step. Assumes first value is eval_loss and
second value is accuracy_top_1.
callbacks: a list of callbacks which might include a time history callback
used during keras.fit.
Returns:
Dictionary of normalized results.
"""
stats = {}
if eval_output:
stats['accuracy_top_1'] = eval_output[1].item()
stats['eval_loss'] = eval_output[0].item()
if history and history.history:
train_hist = history.history
# Gets final loss from training.
stats['loss'] = train_hist['loss'][-1].item()
# Gets top_1 training accuracy.
if 'categorical_accuracy' in train_hist:
stats[TRAIN_TOP_1] = train_hist['categorical_accuracy'][-1].item()
elif 'sparse_categorical_accuracy' in train_hist:
stats[TRAIN_TOP_1] = train_hist['sparse_categorical_accuracy'][-1].item()
if not callbacks:
return stats
# Look for the time history callback which was used during keras.fit
for callback in callbacks:
if isinstance(callback, keras_utils.TimeHistory):
timestamp_log = callback.timestamp_log
stats['step_timestamp_log'] = timestamp_log
stats['train_finish_time'] = callback.train_finish_time
if len(timestamp_log) > 1:
stats['avg_exp_per_second'] = (
callback.batch_size * callback.log_steps *
(len(callback.timestamp_log)-1) /
(timestamp_log[-1].timestamp - timestamp_log[0].timestamp))
return stats
def define_keras_flags(dynamic_loss_scale=True):
"""Define flags for Keras models."""
flags_core.define_base(run_eagerly=True)
flags_core.define_performance(num_parallel_calls=False,
synthetic_data=True,
dtype=True,
all_reduce_alg=True,
num_packs=True,
tf_gpu_thread_mode=True,
datasets_num_private_threads=True,
dynamic_loss_scale=dynamic_loss_scale,
loss_scale=True,
tf_data_experimental_slack=True,
enable_xla=True,
force_v2_in_keras_compile=True)
flags_core.define_image()
flags_core.define_benchmark()
flags_core.define_distribution()
flags.adopt_module_key_flags(flags_core)
flags.DEFINE_boolean(name='enable_eager', default=False, help='Enable eager?')
flags.DEFINE_boolean(name='skip_eval', default=False, help='Skip evaluation?')
# TODO(b/135607288): Remove this flag once we understand the root cause of
# slowdown when setting the learning phase in Keras backend.
flags.DEFINE_boolean(
name='set_learning_phase_to_train', default=True,
help='If skip eval, also set Keras learning phase to 1 (training).')
flags.DEFINE_boolean(
name='explicit_gpu_placement', default=False,
help='If not using distribution strategy, explicitly set device scope '
'for the Keras training loop.')
flags.DEFINE_boolean(name='use_trivial_model', default=False,
help='Whether to use a trivial Keras model.')
flags.DEFINE_boolean(name='report_accuracy_metrics', default=True,
help='Report metrics during training and evaluation.')
flags.DEFINE_boolean(name='use_tensor_lr', default=False,
help='Use learning rate tensor instead of a callback.')
flags.DEFINE_boolean(
name='enable_tensorboard', default=False,
help='Whether to enable Tensorboard callback.')
flags.DEFINE_integer(
name='train_steps', default=None,
help='The number of steps to run for training. If it is larger than '
'# batches per epoch, then use # batches per epoch. When this flag is '
'set, only one epoch is going to run for training.')
flags.DEFINE_string(
name='profile_steps', default=None,
help='Save profiling data to model dir at given range of steps. The '
'value must be a comma separated pair of positive integers, specifying '
'the first and last step to profile. For example, "--profile_steps=2,4" '
'triggers the profiler to process 3 steps, starting from the 2nd step. '
'Note that profiler has a non-trivial performance overhead, and the '
'output file can be gigantic if profiling many steps.')
flags.DEFINE_boolean(
name='data_delay_prefetch', default=False,
help='Add a small delay in tf.data prefetch to prioritize memory copy of '
'other tensors over the data minibatch for the (T+1)th step. It should '
'help improve performance using EagerIterator and function. The codepath '
'when enabling this feature is experimental and will be removed once the '
'corresponding performance features are fully supported in TensorFlow.')
flags.DEFINE_boolean(
name='batchnorm_spatial_persistent', default=True,
help='Enable the spacial persistent mode for CuDNN batch norm kernel.')
flags.DEFINE_boolean(
name='enable_get_next_as_optional', default=False,
help='Enable get_next_as_optional behavior in DistributedIterator.')
def get_synth_input_fn(height, width, num_channels, num_classes,
dtype=tf.float32, drop_remainder=True):
"""Returns an input function that returns a dataset with random data.
This input_fn returns a data set that iterates over a set of random data and
bypasses all preprocessing, e.g. jpeg decode and copy. The host to device
copy is still included. This used to find the upper throughput bound when
tuning the full input pipeline.
Args:
height: Integer height that will be used to create a fake image tensor.
width: Integer width that will be used to create a fake image tensor.
num_channels: Integer depth that will be used to create a fake image tensor.
num_classes: Number of classes that should be represented in the fake labels
tensor
dtype: Data type for features/images.
drop_remainder: A boolean indicates whether to drop the remainder of the
batches. If True, the batch dimension will be static.
Returns:
An input_fn that can be used in place of a real one to return a dataset
that can be used for iteration.
"""
# pylint: disable=unused-argument
def input_fn(is_training, data_dir, batch_size, *args, **kwargs):
"""Returns dataset filled with random data."""
# Synthetic input should be within [0, 255].
inputs = tf.random.truncated_normal([height, width, num_channels],
dtype=dtype,
mean=127,
stddev=60,
name='synthetic_inputs')
labels = tf.random.uniform([1],
minval=0,
maxval=num_classes - 1,
dtype=tf.int32,
name='synthetic_labels')
# Cast to float32 for Keras model.
labels = tf.cast(labels, dtype=tf.float32)
data = tf.data.Dataset.from_tensors((inputs, labels)).repeat()
# `drop_remainder` will make dataset produce outputs with known shapes.
data = data.batch(batch_size, drop_remainder=drop_remainder)
data = data.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
return data
return input_fn
def data_delay_prefetch():
"""Use unstable code for perf tuning purposes."""
if not FLAGS.use_synthetic_data:
_monkey_patch_org_create_device_dataset()
def set_cudnn_batchnorm_mode():
"""Set CuDNN batchnorm mode for better performance.
Note: Spatial Persistent mode may lead to accuracy losses for certain
models.
"""
if FLAGS.batchnorm_spatial_persistent:
os.environ['TF_USE_CUDNN_BATCHNORM_SPATIAL_PERSISTENT'] = '1'
else:
os.environ.pop('TF_USE_CUDNN_BATCHNORM_SPATIAL_PERSISTENT', None)
# TODO(haoyuzhang): remove this monkey patch when the "prefetch with slack"
# feature is available in tf.data.
def _monkey_patch_org_create_device_dataset():
"""Monkey-patch `_create_device_dataset` method with delayed prefetch."""
import ast # pylint: disable=g-import-not-at-top
import inspect # pylint: disable=g-import-not-at-top
from tensorflow.python.data.ops import multi_device_iterator_ops # pylint: disable=g-import-not-at-top
tf.compat.v1.logging.info(
'Using monkey-patched version of MultiDeviceIterator. It should be '
'removed when the prefetch with slack feature is implemented in tf.data.')
cls_multi_device_iterator = ast.parse(
inspect.getsource(multi_device_iterator_ops.MultiDeviceIterator))
org_create_device_dataset_code = inspect.getsource(
multi_device_iterator_ops.MultiDeviceIterator._create_device_dataset) # pylint: disable=protected-access
code_lines = org_create_device_dataset_code.split('\n')
# Insert in reverse order to avoid line number shift by previous insertions
code_lines.insert(5, ' ds = ds.apply(sleep_ops.sleep(11000))') # 11ms
code_lines.insert(2, ' from tensorflow.python.data.experimental.ops import sleep as sleep_ops') # pylint: disable=line-too-long
patched_code = '\n'.join(line[2:] for line in code_lines)
cls_multi_device_iterator.body[0].body[2] = ast.parse(patched_code).body[0]
exec(compile(cls_multi_device_iterator, '<string>', 'exec'), # pylint: disable=exec-used
multi_device_iterator_ops.__dict__)