common.py 15.8 KB
Newer Older
Yeqing Li's avatar
Yeqing Li committed
1
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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.
Yeqing Li's avatar
Yeqing Li committed
14

15
"""Common util functions and classes used by both keras cifar and imagenet."""
16
17
18
19
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

20
import os
21

Toby Boyd's avatar
Toby Boyd committed
22
23
from absl import flags
import tensorflow as tf
24

25
import tensorflow_model_optimization as tfmot
26
from official.utils.flags import core as flags_core
27
from official.utils.misc import keras_utils
28

Shining Sun's avatar
Shining Sun committed
29
FLAGS = flags.FLAGS
Shining Sun's avatar
Shining Sun committed
30
BASE_LEARNING_RATE = 0.1  # This matches Jing's version.
31
TRAIN_TOP_1 = 'training_accuracy_top_1'
Hongkun Yu's avatar
Hongkun Yu committed
32
LR_SCHEDULE = [  # (multiplier, epoch to start) tuples
Hongkun Yu's avatar
Hongkun Yu committed
33
34
35
36
    (1.0, 5), (0.1, 30), (0.01, 60), (0.001, 80)
]


37
38
39
40
class PiecewiseConstantDecayWithWarmup(
    tf.keras.optimizers.schedules.LearningRateSchedule):
  """Piecewise constant decay with warmup schedule."""

Hongkun Yu's avatar
Hongkun Yu committed
41
42
43
44
45
46
47
48
  def __init__(self,
               batch_size,
               epoch_size,
               warmup_epochs,
               boundaries,
               multipliers,
               compute_lr_on_cpu=True,
               name=None):
49
50
51
52
53
54
    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
Zongwei Zhou's avatar
Zongwei Zhou committed
55
    steps_per_epoch = epoch_size // batch_size
56
57

    self.rescaled_lr = BASE_LEARNING_RATE * batch_size / base_lr_batch_size
Zongwei Zhou's avatar
Zongwei Zhou committed
58
    self.step_boundaries = [float(steps_per_epoch) * x for x in boundaries]
59
    self.lr_values = [self.rescaled_lr * m for m in multipliers]
Zongwei Zhou's avatar
Zongwei Zhou committed
60
    self.warmup_steps = warmup_epochs * steps_per_epoch
61
62
63
    self.compute_lr_on_cpu = compute_lr_on_cpu
    self.name = name

64
    self.learning_rate_ops_cache = {}
65
66
67
68
69
70
71
72

  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.
73
74
    graph = tf.compat.v1.get_default_graph()
    if graph not in self.learning_rate_ops_cache:
75
76
      if self.compute_lr_on_cpu:
        with tf.device('/device:CPU:0'):
77
          self.learning_rate_ops_cache[graph] = self._get_learning_rate(step)
78
      else:
79
80
        self.learning_rate_ops_cache[graph] = self._get_learning_rate(step)
    return self.learning_rate_ops_cache[graph]
81
82
83

  def _get_learning_rate(self, step):
    """Compute learning rate at given step."""
ayushmankumar7's avatar
ayushmankumar7 committed
84
    with tf.name_scope('PiecewiseConstantDecayWithWarmup'):
Hongkun Yu's avatar
Hongkun Yu committed
85

86
87
88
      def warmup_lr(step):
        return self.rescaled_lr * (
            tf.cast(step, tf.float32) / tf.cast(self.warmup_steps, tf.float32))
Hongkun Yu's avatar
Hongkun Yu committed
89

90
      def piecewise_lr(step):
Hongkun Yu's avatar
Hongkun Yu committed
91
92
93
94
        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),
95
96
97
98
99
100
101
102
103
104
105
106
107
108
                     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 get_optimizer(learning_rate=0.1):
109
110
  """Returns optimizer to use."""
  # The learning_rate is overwritten at the beginning of each step by callback.
Scott Zhu's avatar
Scott Zhu committed
111
  return tf.keras.optimizers.SGD(learning_rate=learning_rate, momentum=0.9)
112
113


Hongkun Yu's avatar
Hongkun Yu committed
114
115
116
def get_callbacks(pruning_method=None,
                  enable_checkpoint_and_export=False,
                  model_dir=None):
117
  """Returns common callbacks."""
118
119
120
121
  time_callback = keras_utils.TimeHistory(
      FLAGS.batch_size,
      FLAGS.log_steps,
      logdir=FLAGS.model_dir if FLAGS.enable_tensorboard else None)
122
123
  callbacks = [time_callback]

124
125
  if FLAGS.enable_tensorboard:
    tensorboard_callback = tf.keras.callbacks.TensorBoard(
Hongkun Yu's avatar
Hongkun Yu committed
126
        log_dir=FLAGS.model_dir, profile_batch=FLAGS.profile_steps)
127
128
    callbacks.append(tensorboard_callback)

129
130
131
132
  is_pruning_enabled = pruning_method is not None
  if is_pruning_enabled:
    callbacks.append(tfmot.sparsity.keras.UpdatePruningStep())
    if model_dir is not None:
Hongkun Yu's avatar
Hongkun Yu committed
133
134
135
      callbacks.append(
          tfmot.sparsity.keras.PruningSummaries(
              log_dir=model_dir, profile_batch=0))
136
137
138
139
140

  if enable_checkpoint_and_export:
    if model_dir is not None:
      ckpt_full_path = os.path.join(model_dir, 'model.ckpt-{epoch:04d}')
      callbacks.append(
Hongkun Yu's avatar
Hongkun Yu committed
141
142
          tf.keras.callbacks.ModelCheckpoint(
              ckpt_full_path, save_weights_only=True))
143
144
145
146
  return callbacks


def build_stats(history, eval_output, callbacks):
147
148
149
150
151
152
153
  """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.
154
155
    callbacks: a list of callbacks which might include a time history callback
      used during keras.fit.
156
157
158
159
160
161

  Returns:
    Dictionary of normalized results.
  """
  stats = {}
  if eval_output:
162
163
    stats['accuracy_top_1'] = float(eval_output[1])
    stats['eval_loss'] = float(eval_output[0])
164
165
166
  if history and history.history:
    train_hist = history.history
    # Gets final loss from training.
167
    stats['loss'] = float(train_hist['loss'][-1])
168
169
    # Gets top_1 training accuracy.
    if 'categorical_accuracy' in train_hist:
170
      stats[TRAIN_TOP_1] = float(train_hist['categorical_accuracy'][-1])
171
    elif 'sparse_categorical_accuracy' in train_hist:
172
      stats[TRAIN_TOP_1] = float(train_hist['sparse_categorical_accuracy'][-1])
Allen Wang's avatar
Allen Wang committed
173
174
    elif 'accuracy' in train_hist:
      stats[TRAIN_TOP_1] = float(train_hist['accuracy'][-1])
175

176
177
178
179
180
181
182
183
184
  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
185
186
187
      if callback.epoch_runtime_log:
        stats['avg_exp_per_second'] = callback.average_examples_per_second

188
189
190
  return stats


Hongkun Yu's avatar
Hongkun Yu committed
191
192
193
194
def define_keras_flags(dynamic_loss_scale=True,
                       model=False,
                       optimizer=False,
                       pretrained_filepath=False):
195
  """Define flags for Keras models."""
Hongkun Yu's avatar
Hongkun Yu committed
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
  flags_core.define_base(
      clean=True,
      num_gpu=True,
      run_eagerly=True,
      train_epochs=True,
      epochs_between_evals=True,
      distribution_strategy=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,
      fp16_implementation=True,
      tf_data_experimental_slack=True,
      enable_xla=True,
      training_dataset_cache=True)
217
218
  flags_core.define_image()
  flags_core.define_benchmark()
219
  flags_core.define_distribution()
220
  flags.adopt_module_key_flags(flags_core)
221

Shining Sun's avatar
Shining Sun committed
222
  flags.DEFINE_boolean(name='enable_eager', default=False, help='Enable eager?')
223
  flags.DEFINE_boolean(name='skip_eval', default=False, help='Skip evaluation?')
224
225
226
  # 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(
Hongkun Yu's avatar
Hongkun Yu committed
227
228
      name='set_learning_phase_to_train',
      default=True,
229
230
      help='If skip eval, also set Keras learning phase to 1 (training).')
  flags.DEFINE_boolean(
Hongkun Yu's avatar
Hongkun Yu committed
231
232
      name='explicit_gpu_placement',
      default=False,
233
234
      help='If not using distribution strategy, explicitly set device scope '
      'for the Keras training loop.')
235
  flags.DEFINE_boolean(
Hongkun Yu's avatar
Hongkun Yu committed
236
237
238
239
240
241
242
243
244
245
246
247
248
249
      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=True,
      help='Use learning rate tensor instead of a callback.')
  flags.DEFINE_boolean(
      name='enable_tensorboard',
      default=False,
250
      help='Whether to enable Tensorboard callback.')
Hongkun Yu's avatar
Hongkun Yu committed
251
  flags.DEFINE_string(
Hongkun Yu's avatar
Hongkun Yu committed
252
253
      name='profile_steps',
      default=None,
Hongkun Yu's avatar
Hongkun Yu committed
254
255
256
257
258
259
      help='Save profiling data to model dir at given range of global 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.')
Shining Sun's avatar
Shining Sun committed
260
  flags.DEFINE_integer(
Hongkun Yu's avatar
Hongkun Yu committed
261
262
      name='train_steps',
      default=None,
263
      help='The number of steps to run for training. If it is larger than '
264
265
      '# batches per epoch, then use # batches per epoch. This flag will be '
      'ignored if train_epochs is set to be larger than 1. ')
266
  flags.DEFINE_boolean(
Hongkun Yu's avatar
Hongkun Yu committed
267
268
      name='batchnorm_spatial_persistent',
      default=True,
269
      help='Enable the spacial persistent mode for CuDNN batch norm kernel.')
270
  flags.DEFINE_boolean(
Hongkun Yu's avatar
Hongkun Yu committed
271
272
      name='enable_get_next_as_optional',
      default=False,
273
      help='Enable get_next_as_optional behavior in DistributedIterator.')
Hongkun Yu's avatar
Hongkun Yu committed
274
  flags.DEFINE_boolean(
Hongkun Yu's avatar
Hongkun Yu committed
275
276
      name='enable_checkpoint_and_export',
      default=False,
Hongkun Yu's avatar
Hongkun Yu committed
277
      help='Whether to enable a checkpoint callback and export the savedmodel.')
Hongkun Yu's avatar
Hongkun Yu committed
278
  flags.DEFINE_string(name='tpu', default='', help='TPU address to connect to.')
Jing Li's avatar
Jing Li committed
279
  flags.DEFINE_integer(
280
      name='steps_per_loop',
281
      default=None,
282
      help='Number of steps per training loop. Only training step happens '
Jing Li's avatar
Jing Li committed
283
284
      'inside the loop. Callbacks will not be called inside. Will be capped at '
      'steps per epoch.')
285
286
287
288
289
290
  flags.DEFINE_boolean(
      name='use_tf_while_loop',
      default=True,
      help='Whether to build a tf.while_loop inside the training loop on the '
      'host. Setting it to True is critical to have peak performance on '
      'TPU.')
Shining Sun's avatar
Shining Sun committed
291

292
293
294
295
  if model:
    flags.DEFINE_string('model', 'resnet50_v1.5',
                        'Name of model preset. (mobilenet, resnet50_v1.5)')
  if optimizer:
Hongkun Yu's avatar
Hongkun Yu committed
296
297
298
    flags.DEFINE_string(
        'optimizer', 'resnet50_default', 'Name of optimizer preset. '
        '(mobilenet_default, resnet50_default)')
Jaehong Kim's avatar
Jaehong Kim committed
299
    # TODO(kimjaehong): Replace as general hyper-params not only for mobilenet.
Hongkun Yu's avatar
Hongkun Yu committed
300
301
302
303
    flags.DEFINE_float(
        'initial_learning_rate_per_sample', 0.00007,
        'Initial value of learning rate per sample for '
        'mobilenet_default.')
Jaehong Kim's avatar
Jaehong Kim committed
304
305
306
307
    flags.DEFINE_float('lr_decay_factor', 0.94,
                       'Learning rate decay factor for mobilenet_default.')
    flags.DEFINE_float('num_epochs_per_decay', 2.5,
                       'Number of epochs per decay for mobilenet_default.')
308
  if pretrained_filepath:
Hongkun Yu's avatar
Hongkun Yu committed
309
    flags.DEFINE_string('pretrained_filepath', '', 'Pretrained file path.')
310

311

Allen Wang's avatar
Allen Wang committed
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
def get_synth_data(height, width, num_channels, num_classes, dtype):
  """Creates a set of synthetic random data.

  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.

  Returns:
    A tuple of tensors representing the inputs and labels.

  """
  # 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')
  return inputs, labels


341
342
def define_pruning_flags():
  """Define flags for pruning methods."""
Hongkun Yu's avatar
Hongkun Yu committed
343
344
345
  flags.DEFINE_string(
      'pruning_method', None, 'Pruning method.'
      'None (no pruning) or polynomial_decay.')
346
347
348
349
  flags.DEFINE_float('pruning_initial_sparsity', 0.0,
                     'Initial sparsity for pruning.')
  flags.DEFINE_float('pruning_final_sparsity', 0.5,
                     'Final sparsity for pruning.')
Hongkun Yu's avatar
Hongkun Yu committed
350
351
352
  flags.DEFINE_integer('pruning_begin_step', 0, 'Begin step for pruning.')
  flags.DEFINE_integer('pruning_end_step', 100000, 'End step for pruning.')
  flags.DEFINE_integer('pruning_frequency', 100, 'Frequency for pruning.')
353
354


355
356
357
def define_clustering_flags():
  """Define flags for clustering methods."""
  flags.DEFINE_string('clustering_method', None,
358
                      'None (no clustering) or selective_clustering '
359
                      '(cluster last three Conv2D layers of the model).')
360

361

Hongkun Yu's avatar
Hongkun Yu committed
362
363
364
365
366
367
def get_synth_input_fn(height,
                       width,
                       num_channels,
                       num_classes,
                       dtype=tf.float32,
                       drop_remainder=True):
Shining Sun's avatar
Shining Sun committed
368
369
370
371
372
  """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
Shining Sun's avatar
Shining Sun committed
373
  tuning the full input pipeline.
Shining Sun's avatar
Shining Sun committed
374
375
376
377
378
379
380
381

  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.
382
383
    drop_remainder: A boolean indicates whether to drop the remainder of the
      batches. If True, the batch dimension will be static.
Shining Sun's avatar
Shining Sun committed
384
385
386
387
388

  Returns:
    An input_fn that can be used in place of a real one to return a dataset
    that can be used for iteration.
  """
Hongkun Yu's avatar
Hongkun Yu committed
389

Shining Sun's avatar
Shining Sun committed
390
391
392
  # pylint: disable=unused-argument
  def input_fn(is_training, data_dir, batch_size, *args, **kwargs):
    """Returns dataset filled with random data."""
Hongkun Yu's avatar
Hongkun Yu committed
393
394
395
396
397
398
    inputs, labels = get_synth_data(
        height=height,
        width=width,
        num_channels=num_channels,
        num_classes=num_classes,
        dtype=dtype)
399
400
    # Cast to float32 for Keras model.
    labels = tf.cast(labels, dtype=tf.float32)
Shining Sun's avatar
Shining Sun committed
401
    data = tf.data.Dataset.from_tensors((inputs, labels)).repeat()
402
403

    # `drop_remainder` will make dataset produce outputs with known shapes.
404
    data = data.batch(batch_size, drop_remainder=drop_remainder)
405
    data = data.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
Shining Sun's avatar
Shining Sun committed
406
407
408
    return data

  return input_fn
Shining Sun's avatar
Shining Sun committed
409
410


411
def set_cudnn_batchnorm_mode():
Toby Boyd's avatar
Toby Boyd committed
412
413
414
415
416
  """Set CuDNN batchnorm mode for better performance.

     Note: Spatial Persistent mode may lead to accuracy losses for certain
     models.
  """
417
418
419
  if FLAGS.batchnorm_spatial_persistent:
    os.environ['TF_USE_CUDNN_BATCHNORM_SPATIAL_PERSISTENT'] = '1'
  else:
420
    os.environ.pop('TF_USE_CUDNN_BATCHNORM_SPATIAL_PERSISTENT', None)