keras_imagenet_main.py 6.45 KB
Newer Older
1
# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
#
# 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.
# ==============================================================================
"""Runs a ResNet model on the ImageNet dataset."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

from absl import app as absl_app
from absl import flags
import tensorflow as tf  # pylint: disable=g-bad-import-order

from official.resnet import imagenet_main
26
from official.resnet import imagenet_preprocessing
27
from official.resnet import resnet_run_loop
28
from official.resnet.keras import keras_common
Shining Sun's avatar
Shining Sun committed
29
from official.resnet.keras import resnet50
30
31
32
33
from official.utils.flags import core as flags_core
from official.utils.logs import logger
from official.utils.misc import distribution_utils

34
35
# import os
# os.environ['TF2_BEHAVIOR'] = 'enabled'
36
37
38
39

LR_SCHEDULE = [    # (multiplier, epoch to start) tuples
    (1.0, 5), (0.1, 30), (0.01, 60), (0.001, 80)
]
Shining Sun's avatar
Shining Sun committed
40

41

42
def learning_rate_schedule(current_epoch, current_batch, batches_per_epoch, batch_size):
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
  """Handles linear scaling rule, gradual warmup, and LR decay.

  The learning rate starts at 0, then it increases linearly per step.
  After 5 epochs we reach the base learning rate (scaled to account
    for batch size).
  After 30, 60 and 80 epochs the learning rate is divided by 10.
  After 90 epochs training stops and the LR is set to 0. This ensures
    that we train for exactly 90 epochs for reproducibility.

  Args:
    current_epoch: integer, current epoch indexed from 0.
    current_batch: integer, current batch in the current epoch, indexed from 0.

  Returns:
    Adjusted learning rate.
  """
Shining Sun's avatar
Shining Sun committed
59
  initial_learning_rate = keras_common.BASE_LEARNING_RATE * batch_size / 256
60
61
62
63
  epoch = current_epoch + float(current_batch) / batches_per_epoch
  warmup_lr_multiplier, warmup_end_epoch = LR_SCHEDULE[0]
  if epoch < warmup_end_epoch:
    # Learning rate increases linearly per step.
64
    return initial_learning_rate * warmup_lr_multiplier * epoch / warmup_end_epoch
65
66
  for mult, start_epoch in LR_SCHEDULE:
    if epoch >= start_epoch:
67
      learning_rate = initial_learning_rate * mult
68
69
70
71
72
73
    else:
      break
  return learning_rate


def parse_record_keras(raw_record, is_training, dtype):
Shining Sun's avatar
Shining Sun committed
74
  """Adjust the shape of label."""
Shining Sun's avatar
Shining Sun committed
75
  image, label = imagenet_main.parse_record(raw_record, is_training, dtype)
Shining Sun's avatar
Shining Sun committed
76

Shining Sun's avatar
Shining Sun committed
77
78
79
  # Subtract one so that labels are in [0, 1000), and cast to float32 for
  # Keras model.
  label = tf.cast(tf.cast(tf.reshape(label, shape=[1]), dtype=tf.int32) - 1,
Shining Sun's avatar
Shining Sun committed
80
      dtype=tf.float32)
81
82
83
84
85
86
87
88
89
90
91
92
  return image, label


def run_imagenet_with_keras(flags_obj):
  """Run ResNet ImageNet training and eval loop using native Keras APIs.

  Args:
    flags_obj: An object containing parsed flag values.

  Raises:
    ValueError: If fp16 is passed as it is not currently supported.
  """
93
94
  if flags_obj.enable_eager:
    tf.enable_eager_execution()
Shining Sun's avatar
Shining Sun committed
95

96
97
98
99
100
101
102
103
104
105
  dtype = flags_core.get_tf_dtype(flags_obj)
  if dtype == 'fp16':
    raise ValueError('dtype fp16 is not supported in Keras. Use the default '
                     'value(fp32).')

  per_device_batch_size = distribution_utils.per_device_batch_size(
      flags_obj.batch_size, flags_core.get_num_gpus(flags_obj))

  # pylint: disable=protected-access
  if flags_obj.use_synthetic_data:
Shining Sun's avatar
Shining Sun committed
106
    input_fn = keras_common.get_synth_input_fn(
107
108
109
110
        height=imagenet_main.DEFAULT_IMAGE_SIZE,
        width=imagenet_main.DEFAULT_IMAGE_SIZE,
        num_channels=imagenet_main.NUM_CHANNELS,
        num_classes=imagenet_main.NUM_CLASSES,
Shining Sun's avatar
Shining Sun committed
111
        dtype=flags_core.get_tf_dtype(flags_obj))
112
  else:
Shining Sun's avatar
Shining Sun committed
113
    input_fn = imagenet_main.input_fn
114

Shining Sun's avatar
Shining Sun committed
115
116
117
118
119
120
  train_input_dataset = input_fn(
        is_training=True,
        data_dir=flags_obj.data_dir,
        batch_size=per_device_batch_size,
        num_epochs=flags_obj.train_epochs,
        parse_record_fn=parse_record_keras)
121

Shining Sun's avatar
Shining Sun committed
122
123
124
125
126
127
  eval_input_dataset = input_fn(
        is_training=False,
        data_dir=flags_obj.data_dir,
        batch_size=per_device_batch_size,
        num_epochs=flags_obj.train_epochs,
        parse_record_fn=parse_record_keras)
128

Shining Sun's avatar
Shining Sun committed
129
  optimizer = keras_common.get_optimizer()
130
131
  strategy = distribution_utils.get_distribution_strategy(
    flags_obj.num_gpus, flags_obj.use_one_device_strategy)
132

133
  model = resnet50.ResNet50(num_classes=imagenet_main.NUM_CLASSES)
Shining Sun's avatar
Shining Sun committed
134

Shining Sun's avatar
Shining Sun committed
135
  model.compile(loss='sparse_categorical_crossentropy',
Shining Sun's avatar
Shining Sun committed
136
                optimizer=optimizer,
Shining Sun's avatar
Shining Sun committed
137
                metrics=['sparse_categorical_accuracy'],
138
                distribute=strategy)
Shining Sun's avatar
Shining Sun committed
139

140
141
  time_callback, tensorboard_callback, lr_callback = keras_common.get_callbacks(
      learning_rate_schedule, imagenet_main.NUM_IMAGES['train'])
142

143
144
  steps_per_epoch = imagenet_main.NUM_IMAGES['train'] // flags_obj.batch_size
  num_eval_steps = (imagenet_main.NUM_IMAGES['validation'] //
145
                  flags_obj.batch_size)
Shining Sun's avatar
Shining Sun committed
146

Shining Sun's avatar
Shining Sun committed
147
148
149
150
151
152
153
  train_steps = imagenet_main.NUM_IMAGES['train'] // flags_obj.batch_size
  train_epochs = flags_obj.train_epochs

  if flags_obj.train_steps:
    train_steps = min(flags_obj.train_steps, train_steps)
    train_epochs = 1

Shining Sun's avatar
bug fix  
Shining Sun committed
154
  history = model.fit(train_input_dataset,
Shining Sun's avatar
Shining Sun committed
155
156
                      epochs=train_epochs,
                      steps_per_epoch=train_steps,
Shining Sun's avatar
bug fix  
Shining Sun committed
157
158
159
160
161
162
163
164
                      callbacks=[
                        time_callback,
                        lr_callback,
                        tensorboard_callback
                      ],
                      validation_steps=num_eval_steps,
                      validation_data=eval_input_dataset,
                      verbose=1)
Shining Sun's avatar
Shining Sun committed
165

166
167
168
169
  if not flags_obj.skip_eval:
    eval_output = model.evaluate(eval_input_dataset,
                                 steps=num_eval_steps,
                                 verbose=1)
Shining Sun's avatar
bug fix  
Shining Sun committed
170
171

  stats = keras_common.analyze_fit_and_eval_result(history, eval_output)
172
173

  return stats
174

Shining Sun's avatar
Shining Sun committed
175

176
177
178
179
180
181
def main(_):
  with logger.benchmark_context(flags.FLAGS):
    run_imagenet_with_keras(flags.FLAGS)


if __name__ == '__main__':
182
  tf.logging.set_verbosity(tf.logging.INFO)
183
  imagenet_main.define_imagenet_flags()
Shining Sun's avatar
Shining Sun committed
184
  keras_common.define_keras_flags()
185
  absl_app.run(main)