resnet.py 7.83 KB
Newer Older
Abdullah Rashwan's avatar
Abdullah Rashwan committed
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
# Copyright 2020 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.
# ==============================================================================
"""Contains definitions of Residual Networks.

Residual networks (ResNets) were proposed in:
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Deep Residual Learning for Image Recognition. arXiv:1512.03385
"""

# Import libraries
import tensorflow as tf
from official.modeling import tf_utils
from official.vision.beta.modeling.layers import nn_blocks

layers = tf.keras.layers

# Specifications for different ResNet variants.
# Each entry specifies block configurations of the particular ResNet variant.
# Each element in the block configuration is in the following format:
# (block_fn, num_filters, block_repeats)
RESNET_SPECS = {
    18: [
        ('residual', 64, 2),
        ('residual', 128, 2),
        ('residual', 256, 2),
        ('residual', 512, 2),
    ],
    34: [
        ('residual', 64, 3),
        ('residual', 128, 4),
        ('residual', 256, 6),
        ('residual', 512, 3),
    ],
    50: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 4),
        ('bottleneck', 256, 6),
        ('bottleneck', 512, 3),
    ],
    101: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 4),
        ('bottleneck', 256, 23),
        ('bottleneck', 512, 3),
    ],
    152: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 8),
        ('bottleneck', 256, 36),
        ('bottleneck', 512, 3),
    ],
    200: [
        ('bottleneck', 64, 3),
        ('bottleneck', 128, 24),
        ('bottleneck', 256, 36),
        ('bottleneck', 512, 3),
    ],
}


@tf.keras.utils.register_keras_serializable(package='Vision')
class ResNet(tf.keras.Model):
  """Class to build ResNet family model."""

  def __init__(self,
               model_id,
               input_specs=layers.InputSpec(shape=[None, None, None, 3]),
               activation='relu',
               use_sync_bn=False,
               norm_momentum=0.99,
               norm_epsilon=0.001,
               kernel_initializer='VarianceScaling',
               kernel_regularizer=None,
               bias_regularizer=None,
               **kwargs):
    """ResNet initialization function.

    Args:
      model_id: `int` depth of ResNet backbone model.
      input_specs: `tf.keras.layers.InputSpec` specs of the input tensor.
      activation: `str` name of the activation function.
      use_sync_bn: if True, use synchronized batch normalization.
      norm_momentum: `float` normalization omentum for the moving average.
      norm_epsilon: `float` small float added to variance to avoid dividing by
        zero.
      kernel_initializer: kernel_initializer for convolutional layers.
      kernel_regularizer: tf.keras.regularizers.Regularizer object for Conv2D.
                          Default to None.
      bias_regularizer: tf.keras.regularizers.Regularizer object for Conv2d.
                        Default to None.
      **kwargs: keyword arguments to be passed.
    """
    self._model_id = model_id
    self._input_specs = input_specs
    self._use_sync_bn = use_sync_bn
    self._activation = activation
    self._norm_momentum = norm_momentum
    self._norm_epsilon = norm_epsilon
    if use_sync_bn:
      self._norm = layers.experimental.SyncBatchNormalization
    else:
      self._norm = layers.BatchNormalization
    self._kernel_initializer = kernel_initializer
    self._kernel_regularizer = kernel_regularizer
    self._bias_regularizer = bias_regularizer

    if tf.keras.backend.image_data_format() == 'channels_last':
      bn_axis = -1
    else:
      bn_axis = 1

    # Build ResNet.
    inputs = tf.keras.Input(shape=input_specs.shape[1:])

    x = layers.Conv2D(
        filters=64, kernel_size=7, strides=2, use_bias=False, padding='same',
        kernel_initializer=self._kernel_initializer,
        kernel_regularizer=self._kernel_regularizer,
        bias_regularizer=self._bias_regularizer)(
            inputs)
    x = self._norm(
        axis=bn_axis, momentum=norm_momentum, epsilon=norm_epsilon)(
            x)
    x = tf_utils.get_activation(activation)(x)
    x = layers.MaxPool2D(pool_size=3, strides=2, padding='same')(x)

    # TODO(xianzhi): keep a list of blocks to make blocks accessible.
    endpoints = {}
    for i, spec in enumerate(RESNET_SPECS[model_id]):
      if spec[0] == 'residual':
        block_fn = nn_blocks.ResidualBlock
      elif spec[0] == 'bottleneck':
        block_fn = nn_blocks.BottleneckBlock
      else:
        raise ValueError('Block fn `{}` is not supported.'.format(spec[0]))
      x = self._block_group(
          inputs=x,
          filters=spec[1],
          strides=(1 if i == 0 else 2),
          block_fn=block_fn,
          block_repeats=spec[2],
          name='block_group_l{}'.format(i + 2))
Abdullah Rashwan's avatar
Abdullah Rashwan committed
155
      endpoints[str(i + 2)] = x
Abdullah Rashwan's avatar
Abdullah Rashwan committed
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

    self._output_specs = {l: endpoints[l].get_shape() for l in endpoints}

    super(ResNet, self).__init__(inputs=inputs, outputs=endpoints, **kwargs)

  def _block_group(self,
                   inputs,
                   filters,
                   strides,
                   block_fn,
                   block_repeats=1,
                   name='block_group'):
    """Creates one group of blocks for the ResNet model.

    Args:
      inputs: `Tensor` of size `[batch, channels, height, width]`.
      filters: `int` number of filters for the first convolution of the layer.
      strides: `int` stride to use for the first convolution of the layer. If
        greater than 1, this layer will downsample the input.
      block_fn: Either `nn_blocks.ResidualBlock` or `nn_blocks.BottleneckBlock`.
      block_repeats: `int` number of blocks contained in the layer.
      name: `str`name for the block.

    Returns:
      The output `Tensor` of the block layer.
    """
    x = block_fn(
        filters=filters,
        strides=strides,
        use_projection=True,
        kernel_initializer=self._kernel_initializer,
        kernel_regularizer=self._kernel_regularizer,
        bias_regularizer=self._bias_regularizer,
        activation=self._activation,
        use_sync_bn=self._use_sync_bn,
        norm_momentum=self._norm_momentum,
        norm_epsilon=self._norm_epsilon)(
            inputs)

    for _ in range(1, block_repeats):
      x = block_fn(
          filters=filters,
          strides=1,
          use_projection=False,
          kernel_initializer=self._kernel_initializer,
          kernel_regularizer=self._kernel_regularizer,
          bias_regularizer=self._bias_regularizer,
          activation=self._activation,
          use_sync_bn=self._use_sync_bn,
          norm_momentum=self._norm_momentum,
          norm_epsilon=self._norm_epsilon)(
              x)

    return tf.identity(x, name=name)

  def get_config(self):
    config_dict = {
        'model_id': self._model_id,
        'activation': self._activation,
        'use_sync_bn': self._use_sync_bn,
        'norm_momentum': self._norm_momentum,
        'norm_epsilon': self._norm_epsilon,
        'kernel_initializer': self._kernel_initializer,
        'kernel_regularizer': self._kernel_regularizer,
        'bias_regularizer': self._bias_regularizer,
    }
    return config_dict

  @classmethod
  def from_config(cls, config, custom_objects=None):
    return cls(**config)

  @property
  def output_specs(self):
    """A dict of {level: TensorShape} pairs for the model output."""
    return self._output_specs