vs-cnn/model_factor.py at master · PreferredAI/vs-cnn · GitHub

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
import os
import tensorflow as tf
import numpy as np
from layers import conv, lrn, max_pool, fc


class FVS_CNN(object):

  def __init__(self, num_classes, num_factors, factor_id_map, factor_layer='fc7',
               skip_layers=None, weights_path='weights/bvlc_alexnet.npy'):
    self.num_classes = num_classes
    self.num_factors = num_factors
    self.factor_layer = factor_layer
    self.factor_weight_size = self._get_weight_size(factor_layer, num_factors)
    self.skip_layers = skip_layers
    self.weights_path = weights_path

    self.factor_id_map = factor_id_map
    self.factor_weight_dict = {}
    self.factor_bias_dict = {}

    self._build_graph()

  def _get_weight_size(self, factor_layer, num_factors):
    if factor_layer == 'conv1':
      return [11, 11, 3, num_factors]
    elif factor_layer == 'conv3':
      return [3, 3, 128, num_factors]
    elif factor_layer == 'conv5':
      return [3, 3, 192, num_factors]
    elif factor_layer == 'fc7':
      return [4096, num_factors]
    else:
      raise ValueError('Invalid factor layer name: {}'.format(factor_layer))

  def _build_graph(self):
    self.x = tf.placeholder(tf.float32, [None, 227, 227, 3])
    self.y = tf.placeholder(tf.float32, [None, 2])

    if self.factor_layer == 'conv1':
      self.factor_weights = tf.placeholder(tf.float32, shape=[11, 11, 3, self.num_factors])
    elif self.factor_layer == 'conv3':
      self.factor_weights = tf.placeholder(tf.float32, shape=[3, 3, 128, self.num_factors])
    elif self.factor_layer == 'conv5':
      self.factor_weights = tf.placeholder(tf.float32, shape=[3, 3, 192, self.num_factors])
    elif self.factor_layer == 'fc7':
      self.factor_weights = tf.placeholder(tf.float32, shape=[4096, self.num_factors])

    self.factor_biases = tf.placeholder(tf.float32, shape=[self.num_factors])

    # 1st Layer: Conv (w ReLu) -> Lrn -> Pool
    if self.factor_layer == 'conv1':
      conv1_factor = conv(self.x, 11, 11, self.num_factors, 4, 4, padding='VALID', name='conv1_factor')
      conv1_shared = conv(self.x, 11, 11, 96 - self.num_factors, 4, 4, padding='VALID', name='conv1_shared')
      splits = tf.split(axis=3, num_or_size_splits=2, value=conv1_factor)
      conv1 = tf.concat([splits[0], conv1_shared, splits[1]], axis=3, name='conv1_concat')
    else:
      conv1 = conv(self.x, 11, 11, 96, 4, 4, padding='VALID', name='conv1')

    norm1 = lrn(conv1, 2, 1e-05, 0.75, name='norm1')
    pool1 = max_pool(norm1, 3, 3, 2, 2, padding='VALID', name='pool1')

    # 2nd Layer: Conv (w ReLu)  -> Lrn -> Pool with 2 groups
    conv2 = conv(pool1, 5, 5, 256, 1, 1, groups=2, name='conv2')
    norm2 = lrn(conv2, 2, 1e-05, 0.75, name='norm2')
    pool2 = max_pool(norm2, 3, 3, 2, 2, padding='VALID', name='pool2')

    # 3rd Layer: Conv (w ReLu)
    if self.factor_layer == 'conv3':
      conv3_factor = conv(pool2, 3, 3, self.num_factors, 1, 1, groups=2, name='conv3_factor')
      conv3_shared = conv(pool2, 3, 3, 384 - self.num_factors, 1, 1, name='conv3_shared')
      splits = tf.split(axis=3, num_or_size_splits=2, value=conv3_factor)
      conv3 = tf.concat([splits[0], conv3_shared, splits[1]], axis=3, name='conv3_concat')
    else:
      conv3 = conv(pool2, 3, 3, 384, 1, 1, name='conv3')

    # 4th Layer: Conv (w ReLu) split into two groups
    conv4 = conv(conv3, 3, 3, 384, 1, 1, groups=2, name='conv4')

    # 5th Layer: Conv (w ReLu) -> Pool split into two groups
    if self.factor_layer == 'conv5':
      conv5_factor = conv(conv4, 3, 3, self.num_factors, 1, 1, groups=2, name='conv5_factor')
      conv5_shared = conv(conv4, 3, 3, 256 - self.num_factors, 1, 1, groups=2, name='conv5_shared')
      splits = tf.split(axis=3, num_or_size_splits=2, value=conv5_factor)
      conv5 = tf.concat([splits[0], conv5_shared, splits[1]], axis=3, name='conv5_concat')
    else:
      conv5 = conv(conv4, 3, 3, 256, 1, 1, groups=2, name='conv5')

    pool5 = max_pool(conv5, 3, 3, 2, 2, padding='VALID', name='pool5')

    # 6th Layer: Flatten -> FC (w ReLu)
    flattened = tf.reshape(pool5, [-1, 6 * 6 * 256])
    fc6 = fc(flattened, 6 * 6 * 256, 4096, name='fc6')

    if self.factor_layer == 'fc7':
      fc7_factor = fc(fc6, 4096, self.num_factors, 'fc7_factor')
      fc7_shared = fc(fc6, 4096, 4096 - self.num_factors, 'fc7_shared')
      fc7 = tf.concat([fc7_factor, fc7_shared], axis=1, name='fc7_concat')
    else:
      fc7 = fc(fc6, 4096, 4096, name='fc7')

    with tf.variable_scope('{}_factor'.format(self.factor_layer), reuse=True):
      self.assign_factor = tf.group(tf.get_variable('weights').assign(self.factor_weights),
                                    tf.get_variable('biases').assign(self.factor_biases))

    # 8th Layer: FC and return unscaled activations
    self.fc8 = fc(fc7, 4096, self.num_classes, relu=False, name='fc8')
    self.prob = tf.nn.softmax(self.fc8, name='prob')

  def load_factor_weights(self, session, factor_ids, weight_dir=None):
    factor_id = factor_ids[0]
    if factor_id in self.factor_weight_dict.keys():
      weights = self.factor_weight_dict[factor_id]
      biases = self.factor_bias_dict[factor_id]
    else:
      factor = self.factor_id_map[factor_id]
      if weight_dir:
        weights = np.load(os.path.join(weight_dir, factor, 'weights.npy'))
        biases = np.load(os.path.join(weight_dir, factor, 'biases.npy'))
      else:
        if self.factor_layer.startswith('conv'):
          weights = np.random.normal(0.0, scale=0.01, size=self.factor_weight_size)
        else:
          weights = np.random.normal(0.0, scale=0.005, size=self.factor_weight_size)
        biases = np.zeros(self.num_factors)
      self.factor_weight_dict[factor_id] = weights
      self.factor_bias_dict[factor_id] = biases
    session.run(self.assign_factor, feed_dict={self.factor_weights: weights,
                                               self.factor_biases: biases})

  def update_factor_weights(self, session, factor_ids):
    factor_id = factor_ids[0]
    with tf.variable_scope('{}_factor'.format(self.factor_layer), reuse=True):
      self.factor_weight_dict[factor_id], self.factor_bias_dict[factor_id] = \
        session.run([tf.get_variable('weights'), tf.get_variable('biases')])

  def load_initial_weights(self, session):
    print('Loading initial weights from:', self.weights_path)
    self.weights_dict = dict(np.load(self.weights_path, encoding='bytes').item())

    for op_name in self.weights_dict.keys():
      if op_name not in self.skip_layers:
        print('Loading weights for layer:', op_name)
        with tf.variable_scope(op_name, reuse=True):
          for data in self.weights_dict[op_name]:
            if len(data.shape) == 1:
              var = tf.get_variable('biases', trainable=True)
              session.run(var.assign(data))
            else:
              var = tf.get_variable('weights', trainable=True)
              session.run(var.assign(data))

    print('Loading weights for layer: {}_shared'.format(self.factor_layer))
    if self.factor_layer.startswith('conv'):
      with tf.variable_scope('{}_shared'.format(self.factor_layer), reuse=True):
        for data in self.weights_dict[self.factor_layer]:
          if len(data.shape) == 1:
            var = tf.get_variable('biases', trainable=True)
            session.run(var.assign(data[int(self.num_factors / 2):-int(self.num_factors / 2)]))
          else:
            var = tf.get_variable('weights', trainable=True)
            session.run(var.assign(data[:, :, :, int(self.num_factors / 2):-int(self.num_factors / 2)]))

    elif self.factor_layer == 'fc7':
      with tf.variable_scope('fc7_shared', reuse=True):
        for data in self.weights_dict['fc7']:
          if len(data.shape) == 1:
            var = tf.get_variable('biases', trainable=True)
            session.run(var.assign(data[self.num_factors:]))
          else:
            var = tf.get_variable('weights', trainable=True)
            session.run(var.assign(data[:, self.num_factors:]))