File size: 9,381 Bytes
0b8359d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Lint as: python2, python3
# 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.
# ==============================================================================

"""Cell structure used by NAS."""

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

import functools
from six.moves import range
from six.moves import zip
import tensorflow as tf
from tensorflow.contrib import framework as contrib_framework
from tensorflow.contrib import slim as contrib_slim
from deeplab.core import xception as xception_utils
from deeplab.core.utils import resize_bilinear
from deeplab.core.utils import scale_dimension
from tensorflow.contrib.slim.nets import resnet_utils

arg_scope = contrib_framework.arg_scope
slim = contrib_slim

separable_conv2d_same = functools.partial(xception_utils.separable_conv2d_same,
                                          regularize_depthwise=True)


class NASBaseCell(object):
  """NASNet Cell class that is used as a 'layer' in image architectures."""

  def __init__(self, num_conv_filters, operations, used_hiddenstates,
               hiddenstate_indices, drop_path_keep_prob, total_num_cells,
               total_training_steps, batch_norm_fn=slim.batch_norm):
    """Init function.

    For more details about NAS cell, see
    https://arxiv.org/abs/1707.07012 and https://arxiv.org/abs/1712.00559.

    Args:
      num_conv_filters: The number of filters for each convolution operation.
      operations: List of operations that are performed in the NASNet Cell in
        order.
      used_hiddenstates: Binary array that signals if the hiddenstate was used
        within the cell. This is used to determine what outputs of the cell
        should be concatenated together.
      hiddenstate_indices: Determines what hiddenstates should be combined
        together with the specified operations to create the NASNet cell.
      drop_path_keep_prob: Float, drop path keep probability.
      total_num_cells: Integer, total number of cells.
      total_training_steps: Integer, total training steps.
      batch_norm_fn: Function, batch norm function. Defaults to
        slim.batch_norm.
    """
    if len(hiddenstate_indices) != len(operations):
      raise ValueError(
          'Number of hiddenstate_indices and operations should be the same.')
    if len(operations) % 2:
      raise ValueError('Number of operations should be even.')
    self._num_conv_filters = num_conv_filters
    self._operations = operations
    self._used_hiddenstates = used_hiddenstates
    self._hiddenstate_indices = hiddenstate_indices
    self._drop_path_keep_prob = drop_path_keep_prob
    self._total_num_cells = total_num_cells
    self._total_training_steps = total_training_steps
    self._batch_norm_fn = batch_norm_fn

  def __call__(self, net, scope, filter_scaling, stride, prev_layer, cell_num):
    """Runs the conv cell."""
    self._cell_num = cell_num
    self._filter_scaling = filter_scaling
    self._filter_size = int(self._num_conv_filters * filter_scaling)

    with tf.variable_scope(scope):
      net = self._cell_base(net, prev_layer)
      for i in range(len(self._operations) // 2):
        with tf.variable_scope('comb_iter_{}'.format(i)):
          h1 = net[self._hiddenstate_indices[i * 2]]
          h2 = net[self._hiddenstate_indices[i * 2 + 1]]
          with tf.variable_scope('left'):
            h1 = self._apply_conv_operation(
                h1, self._operations[i * 2], stride,
                self._hiddenstate_indices[i * 2] < 2)
          with tf.variable_scope('right'):
            h2 = self._apply_conv_operation(
                h2, self._operations[i * 2 + 1], stride,
                self._hiddenstate_indices[i * 2 + 1] < 2)
          with tf.variable_scope('combine'):
            h = h1 + h2
          net.append(h)

      with tf.variable_scope('cell_output'):
        net = self._combine_unused_states(net)

      return net

  def _cell_base(self, net, prev_layer):
    """Runs the beginning of the conv cell before the chosen ops are run."""
    filter_size = self._filter_size

    if prev_layer is None:
      prev_layer = net
    else:
      if net.shape[2] != prev_layer.shape[2]:
        prev_layer = resize_bilinear(
            prev_layer, tf.shape(net)[1:3], prev_layer.dtype)
      if filter_size != prev_layer.shape[3]:
        prev_layer = tf.nn.relu(prev_layer)
        prev_layer = slim.conv2d(prev_layer, filter_size, 1, scope='prev_1x1')
        prev_layer = self._batch_norm_fn(prev_layer, scope='prev_bn')

    net = tf.nn.relu(net)
    net = slim.conv2d(net, filter_size, 1, scope='1x1')
    net = self._batch_norm_fn(net, scope='beginning_bn')
    net = tf.split(axis=3, num_or_size_splits=1, value=net)
    net.append(prev_layer)
    return net

  def _apply_conv_operation(self, net, operation, stride,
                            is_from_original_input):
    """Applies the predicted conv operation to net."""
    if stride > 1 and not is_from_original_input:
      stride = 1
    input_filters = net.shape[3]
    filter_size = self._filter_size
    if 'separable' in operation:
      num_layers = int(operation.split('_')[-1])
      kernel_size = int(operation.split('x')[0][-1])
      for layer_num in range(num_layers):
        net = tf.nn.relu(net)
        net = separable_conv2d_same(
            net,
            filter_size,
            kernel_size,
            depth_multiplier=1,
            scope='separable_{0}x{0}_{1}'.format(kernel_size, layer_num + 1),
            stride=stride)
        net = self._batch_norm_fn(
            net, scope='bn_sep_{0}x{0}_{1}'.format(kernel_size, layer_num + 1))
        stride = 1
    elif 'atrous' in operation:
      kernel_size = int(operation.split('x')[0][-1])
      net = tf.nn.relu(net)
      if stride == 2:
        scaled_height = scale_dimension(tf.shape(net)[1], 0.5)
        scaled_width = scale_dimension(tf.shape(net)[2], 0.5)
        net = resize_bilinear(net, [scaled_height, scaled_width], net.dtype)
        net = resnet_utils.conv2d_same(
            net, filter_size, kernel_size, rate=1, stride=1,
            scope='atrous_{0}x{0}'.format(kernel_size))
      else:
        net = resnet_utils.conv2d_same(
            net, filter_size, kernel_size, rate=2, stride=1,
            scope='atrous_{0}x{0}'.format(kernel_size))
      net = self._batch_norm_fn(net, scope='bn_atr_{0}x{0}'.format(kernel_size))
    elif operation in ['none']:
      if stride > 1 or (input_filters != filter_size):
        net = tf.nn.relu(net)
        net = slim.conv2d(net, filter_size, 1, stride=stride, scope='1x1')
        net = self._batch_norm_fn(net, scope='bn_1')
    elif 'pool' in operation:
      pooling_type = operation.split('_')[0]
      pooling_shape = int(operation.split('_')[-1].split('x')[0])
      if pooling_type == 'avg':
        net = slim.avg_pool2d(net, pooling_shape, stride=stride, padding='SAME')
      elif pooling_type == 'max':
        net = slim.max_pool2d(net, pooling_shape, stride=stride, padding='SAME')
      else:
        raise ValueError('Unimplemented pooling type: ', pooling_type)
      if input_filters != filter_size:
        net = slim.conv2d(net, filter_size, 1, stride=1, scope='1x1')
        net = self._batch_norm_fn(net, scope='bn_1')
    else:
      raise ValueError('Unimplemented operation', operation)

    if operation != 'none':
      net = self._apply_drop_path(net)
    return net

  def _combine_unused_states(self, net):
    """Concatenates the unused hidden states of the cell."""
    used_hiddenstates = self._used_hiddenstates
    states_to_combine = ([
        h for h, is_used in zip(net, used_hiddenstates) if not is_used])
    net = tf.concat(values=states_to_combine, axis=3)
    return net

  @contrib_framework.add_arg_scope
  def _apply_drop_path(self, net):
    """Apply drop_path regularization."""
    drop_path_keep_prob = self._drop_path_keep_prob
    if drop_path_keep_prob < 1.0:
      # Scale keep prob by layer number.
      assert self._cell_num != -1
      layer_ratio = (self._cell_num + 1) / float(self._total_num_cells)
      drop_path_keep_prob = 1 - layer_ratio * (1 - drop_path_keep_prob)
      # Decrease keep prob over time.
      current_step = tf.cast(tf.train.get_or_create_global_step(), tf.float32)
      current_ratio = tf.minimum(1.0, current_step / self._total_training_steps)
      drop_path_keep_prob = (1 - current_ratio * (1 - drop_path_keep_prob))
      # Drop path.
      noise_shape = [tf.shape(net)[0], 1, 1, 1]
      random_tensor = drop_path_keep_prob
      random_tensor += tf.random_uniform(noise_shape, dtype=tf.float32)
      binary_tensor = tf.cast(tf.floor(random_tensor), net.dtype)
      keep_prob_inv = tf.cast(1.0 / drop_path_keep_prob, net.dtype)
      net = net * keep_prob_inv * binary_tensor
    return net