imagenet_main.py 8.92 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# Copyright 2017 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.
# ==============================================================================
15
"""Runs a ResNet model on the ImageNet dataset."""
16
17
18
19
20
21
22

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

import argparse
import os
23
import sys
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

import tensorflow as tf

import resnet_model
import vgg_preprocessing

parser = argparse.ArgumentParser()

parser.add_argument(
    '--data_dir', type=str, default='',
    help='The directory where the ImageNet input data is stored.')

parser.add_argument(
    '--model_dir', type=str, default='/tmp/resnet_model',
    help='The directory where the model will be stored.')

parser.add_argument(
    '--resnet_size', type=int, default=50, choices=[18, 34, 50, 101, 152, 200],
    help='The size of the ResNet model to use.')

parser.add_argument(
45
46
    '--train_epochs', type=int, default=100,
    help='The number of epochs to use for training.')
47
48

parser.add_argument(
49
50
    '--epochs_per_eval', type=int, default=1,
    help='The number of training epochs to run between evaluations.')
51
52

parser.add_argument(
53
54
    '--batch_size', type=int, default=32,
    help='Batch size for training and evaluation.')
55

56
57
58
59
60
61
62
63
parser.add_argument(
    '--data_format', type=str, default=None,
    choices=['channels_first', 'channels_last'],
    help='A flag to override the data format used in the model. channels_first '
         'provides a performance boost on GPU but is not always compatible '
         'with CPU. If left unspecified, the data format will be chosen '
         'automatically based on whether TensorFlow was built for CPU or GPU.')

64
_DEFAULT_IMAGE_SIZE = 224
65
66
_NUM_CHANNELS = 3
_LABEL_CLASSES = 1001
67
68
69
70

_MOMENTUM = 0.9
_WEIGHT_DECAY = 1e-4

71
72
73
74
_NUM_IMAGES = {
    'train': 1281167,
    'validation': 50000,
}
75

76
77
78
79
80
81

def filenames(is_training):
  """Return filenames for dataset."""
  if is_training:
    return [
        os.path.join(FLAGS.data_dir, 'train-%05d-of-01024' % i)
Amit Patankar's avatar
Amit Patankar committed
82
        for i in range(0, 1024)]
83
84
85
  else:
    return [
        os.path.join(FLAGS.data_dir, 'validation-%05d-of-00128' % i)
Amit Patankar's avatar
Amit Patankar committed
86
        for i in range(0, 128)]
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


def dataset_parser(value, is_training):
  """Parse an Imagenet record from value."""
  keys_to_features = {
      'image/encoded':
          tf.FixedLenFeature((), tf.string, default_value=''),
      'image/format':
          tf.FixedLenFeature((), tf.string, default_value='jpeg'),
      'image/class/label':
          tf.FixedLenFeature([], dtype=tf.int64, default_value=-1),
      'image/class/text':
          tf.FixedLenFeature([], dtype=tf.string, default_value=''),
      'image/object/bbox/xmin':
          tf.VarLenFeature(dtype=tf.float32),
      'image/object/bbox/ymin':
          tf.VarLenFeature(dtype=tf.float32),
      'image/object/bbox/xmax':
          tf.VarLenFeature(dtype=tf.float32),
      'image/object/bbox/ymax':
          tf.VarLenFeature(dtype=tf.float32),
      'image/object/class/label':
          tf.VarLenFeature(dtype=tf.int64),
  }

  parsed = tf.parse_single_example(value, keys_to_features)
113

114
115
116
117
118
  image = tf.image.decode_image(
      tf.reshape(parsed['image/encoded'], shape=[]),
      _NUM_CHANNELS)
  image = tf.image.convert_image_dtype(image, dtype=tf.float32)

119
  image = vgg_preprocessing.preprocess_image(
120
      image=image,
121
122
      output_height=_DEFAULT_IMAGE_SIZE,
      output_width=_DEFAULT_IMAGE_SIZE,
123
124
125
126
127
128
129
      is_training=is_training)

  label = tf.cast(
      tf.reshape(parsed['image/class/label'], shape=[]),
      dtype=tf.int32)

  return image, tf.one_hot(label, _LABEL_CLASSES)
130
131


132
133
def input_fn(is_training, num_epochs=1):
  """Input function which provides batches for train or eval."""
134
135
136
137
138
139
  dataset = tf.contrib.data.Dataset.from_tensor_slices(filenames(is_training))
  if is_training:
    dataset = dataset.shuffle(buffer_size=1024)
  dataset = dataset.flat_map(tf.contrib.data.TFRecordDataset)

  if is_training:
140
    dataset = dataset.repeat(num_epochs)
141
142

  dataset = dataset.map(lambda value: dataset_parser(value, is_training),
143
                        num_threads=5,
144
145
146
147
148
149
150
151
                        output_buffer_size=FLAGS.batch_size)

  if is_training:
    buffer_size = 1250 + 2 * FLAGS.batch_size
    dataset = dataset.shuffle(buffer_size=buffer_size)

  iterator = dataset.batch(FLAGS.batch_size).make_one_shot_iterator()
  images, labels = iterator.get_next()
152
153
154
155
  return images, labels


def resnet_model_fn(features, labels, mode):
156
  """Our model_fn for ResNet to be used with our Estimator."""
157
158
  tf.summary.image('images', features, max_outputs=6)

159
160
  network = resnet_model.imagenet_resnet_v2(
      FLAGS.resnet_size, _LABEL_CLASSES, FLAGS.data_format)
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
  logits = network(
      inputs=features, is_training=(mode == tf.estimator.ModeKeys.TRAIN))

  predictions = {
      'classes': tf.argmax(logits, axis=1),
      'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
  }

  if mode == tf.estimator.ModeKeys.PREDICT:
    return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)

  # Calculate loss, which includes softmax cross entropy and L2 regularization.
  cross_entropy = tf.losses.softmax_cross_entropy(
      logits=logits, onehot_labels=labels)

  # Create a tensor named cross_entropy for logging purposes.
  tf.identity(cross_entropy, name='cross_entropy')
  tf.summary.scalar('cross_entropy', cross_entropy)

  # Add weight decay to the loss. We perform weight decay on all trainable
  # variables, which includes batch norm beta and gamma variables.
  loss = cross_entropy + _WEIGHT_DECAY * tf.add_n(
      [tf.nn.l2_loss(v) for v in tf.trainable_variables()])

  if mode == tf.estimator.ModeKeys.TRAIN:
186
187
188
189
    # Scale the learning rate linearly with the batch size. When the batch size is
    # 256, the learning rate should be 0.1.
    initial_learning_rate = 0.1 * FLAGS.batch_size / 256
    batches_per_epoch = _NUM_IMAGES['train'] / FLAGS.batch_size
190
191
    global_step = tf.train.get_or_create_global_step()

192
    # Multiply the learning rate by 0.1 at 30, 60, 80, and 90 epochs.
193
    boundaries = [
194
        int(batches_per_epoch * epoch) for epoch in [30, 60, 80, 90]]
195
    values = [
196
        initial_learning_rate * decay for decay in [1, 0.1, 0.01, 1e-3, 1e-4]]
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
    learning_rate = tf.train.piecewise_constant(
        tf.cast(global_step, tf.int32), boundaries, values)

    # Create a tensor named learning_rate for logging purposes.
    tf.identity(learning_rate, name='learning_rate')
    tf.summary.scalar('learning_rate', learning_rate)

    optimizer = tf.train.MomentumOptimizer(
        learning_rate=learning_rate,
        momentum=_MOMENTUM)

    # Batch norm requires update_ops to be added as a train_op dependency.
    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
    with tf.control_dependencies(update_ops):
      train_op = optimizer.minimize(loss, global_step)
  else:
    train_op = None

  accuracy = tf.metrics.accuracy(
      tf.argmax(labels, axis=1), predictions['classes'])
  metrics = {'accuracy': accuracy}

  # Create a tensor named train_accuracy for logging purposes.
  tf.identity(accuracy[1], name='train_accuracy')
  tf.summary.scalar('train_accuracy', accuracy[1])

  return tf.estimator.EstimatorSpec(
      mode=mode,
      predictions=predictions,
      loss=loss,
      train_op=train_op,
      eval_metric_ops=metrics)


def main(unused_argv):
  # Using the Winograd non-fused algorithms provides a small performance boost.
  os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'

235
236
  # Set up a RunConfig to only save checkpoints once per training cycle.
  run_config = tf.estimator.RunConfig().replace(save_checkpoints_secs=1e9)
237
  resnet_classifier = tf.estimator.Estimator(
238
      model_fn=resnet_model_fn, model_dir=FLAGS.model_dir, config=run_config)
239

240
  for _ in range(FLAGS.train_epochs // FLAGS.epochs_per_eval):
241
242
243
244
245
246
247
248
249
250
251
    tensors_to_log = {
        'learning_rate': 'learning_rate',
        'cross_entropy': 'cross_entropy',
        'train_accuracy': 'train_accuracy'
    }

    logging_hook = tf.train.LoggingTensorHook(
        tensors=tensors_to_log, every_n_iter=100)

    print('Starting a training cycle.')
    resnet_classifier.train(
252
253
        input_fn=lambda: input_fn(
            is_training=True, num_epochs=FLAGS.epochs_per_eval),
254
255
256
        hooks=[logging_hook])

    print('Starting to evaluate.')
257
258
    eval_results = resnet_classifier.evaluate(
        input_fn=lambda: input_fn(is_training=False))
259
260
261
262
263
    print(eval_results)


if __name__ == '__main__':
  tf.logging.set_verbosity(tf.logging.INFO)
264
265
  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(argv=[sys.argv[0]] + unparsed)