-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
167 lines (140 loc) · 8.78 KB
/
model.py
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
'''
referred from:
http://tinyclouds.org/colorize/
http://warmspringwinds.github.io/tensorflow/tf-slim/2016/11/22/upsampling-and-image-segmentation-with-tensorflow-and-tf-slim/
'''
import tensorflow as tf
import numpy as np
class Model:
'''
This is the class for the CNN model and its parameters and hyperparameters
'''
# weights of the network, initialization is glorot (Xavier) initialization
weights = {
'conv0_1_W': tf.get_variable('conv0_1_W', shape=[3,3,1,3], dtype=tf.float32),
'conv1_1_W': tf.get_variable('conv1_1_W', shape=[3,3,3,64], dtype=tf.float32),
'conv1_2_W': tf.get_variable('conv1_2_W', shape=[3,3,64,64], dtype=tf.float32),
'conv2_1_W': tf.get_variable('conv2_1_W', shape=[3,3,64,128], dtype=tf.float32),
'conv2_2_W': tf.get_variable('conv2_2_W', shape=[3,3,128,128], dtype=tf.float32),
'conv3_1_W': tf.get_variable('conv3_1_W', shape=[3,3,128,192], dtype=tf.float32),
'conv3_2_W': tf.get_variable('conv3_2_W', shape=[3,3,192,192], dtype=tf.float32),
'conv3_3_W': tf.get_variable('conv3_3_W', shape=[3,3,192,192], dtype=tf.float32),
'conv4_1_W': tf.get_variable('conv4_1_W', shape=[3,3,192,256], dtype=tf.float32),
'conv4_2_W': tf.get_variable('conv4_2_W', shape=[3,3,256,256], dtype=tf.float32),
'conv4_3_W': tf.get_variable('conv4_3_W', shape=[3,3,256,256], dtype=tf.float32),
'temp_conv1_W': tf.get_variable('temp_conv1_W', shape=[1,1,256,192], dtype=tf.float32),
'temp_conv2_W': tf.get_variable('temp_conv2_W', shape=[3,3,192,128], dtype=tf.float32),
'temp_conv3_W': tf.get_variable('temp_conv3_W', shape=[3,3,128,64], dtype=tf.float32),
'temp_conv4_W': tf.get_variable('temp_conv4_W', shape=[3,3,64,3], dtype=tf.float32),
'temp_conv5_W': tf.get_variable('temp_conv5_W', shape=[3,3,3,3], dtype=tf.float32),
'uv_conv_W': tf.get_variable('uv_conv_W', shape=[3,3,3,2], dtype=tf.float32),
}
# biases of the network, initialization is glorot (Xavier) initialization
biases = {
'conv0_1_b': tf.get_variable('conv0_1_b', shape=[3], dtype=tf.float32),
'conv1_1_b': tf.get_variable('conv1_1_b', shape=[64], dtype=tf.float32),
'conv1_2_b': tf.get_variable('conv1_2_b', shape=[64], dtype=tf.float32),
'conv2_1_b': tf.get_variable('conv2_1_b', shape=[128], dtype=tf.float32),
'conv2_2_b': tf.get_variable('conv2_2_b', shape=[128], dtype=tf.float32),
'conv3_1_b': tf.get_variable('conv3_1_b', shape=[192], dtype=tf.float32),
'conv3_2_b': tf.get_variable('conv3_2_b', shape=[192], dtype=tf.float32),
'conv3_3_b': tf.get_variable('conv3_3_b', shape=[192], dtype=tf.float32),
'conv4_1_b': tf.get_variable('conv4_1_b', shape=[256], dtype=tf.float32),
'conv4_2_b': tf.get_variable('conv4_2_b', shape=[256], dtype=tf.float32),
'conv4_3_b': tf.get_variable('conv4_3_b', shape=[256], dtype=tf.float32),
'temp_conv1_b': tf.get_variable('temp_conv1_b', shape=[192], dtype=tf.float32),
'temp_conv2_b': tf.get_variable('temp_conv2_b', shape=[128], dtype=tf.float32),
'temp_conv3_b': tf.get_variable('temp_conv3_b', shape=[64], dtype=tf.float32),
'temp_conv4_b': tf.get_variable('temp_conv4_b', shape=[3], dtype=tf.float32),
'temp_conv5_b': tf.get_variable('temp_conv5_b', shape=[3], dtype=tf.float32),
'uv_conv_b': tf.get_variable('uv_conv_b', shape=[2], dtype=tf.float32),
}
def __init__(self, batch_size, dim_1, dim_2, num_input_channels=1, num_output_channels=2):
'''Constructor for the Model class.
Arguments:
self {Model} -- instance of the class
batch_size {int} -- batch size for training and testing
dim1 {int} -- dimension 1 (width) of the resize
dim2 {int} -- dimension 2 (height) of the resize
Keyword Arguments:
num_input_channels {int} -- number of channels in the input dataset (default: {1})
num_output_channels {int} -- number of channels in the output dataset (default: {2})
'''
Model.batch_size = batch_size,
Model.dim_1 = dim_1
Model.dim_2 = dim_2
Model.num_input_channels = num_input_channels
Model.num_output_channels = num_output_channels
def CNN_architecture(self, x):
'''the network architecture implementation
Arguments:
self {Model} -- instance of the class
x {tf.placeholder} -- placeholder for input. It must have shape as [None, dim1, dim2, num_input_channels]
Returns:
uv_output {tf.tensor} -- model output after forward pass. It will have shape [None, dim1, dim2, num_output_channels]
'''
conv0_1 = tf.nn.conv2d(x, Model.weights['conv0_1_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv0_1_b']
# 112,112
conv1_1 = tf.nn.conv2d(conv0_1, Model.weights['conv1_1_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv1_1_b']
conv1_1 = tf.contrib.layers.batch_norm(conv1_1)
relu1_1 = tf.nn.relu(conv1_1)
conv1_2 = tf.nn.conv2d(relu1_1, Model.weights['conv1_2_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv1_2_b']
relu1_2 = tf.nn.relu(conv1_2)
# 112,112
hyper1 = tf.contrib.layers.batch_norm(relu1_2)
# 112,112
pool1 = tf.nn.max_pool(hyper1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
# 56,56
conv2_1 = tf.nn.conv2d(pool1, Model.weights['conv2_1_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv2_1_b']
conv2_1 = tf.contrib.layers.batch_norm(conv2_1)
relu2_1 = tf.nn.relu(conv2_1)
conv2_2 = tf.nn.conv2d(relu2_1, Model.weights['conv2_2_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv2_2_b']
relu2_2 = tf.nn.relu(conv2_2)
# 56,56
hyper2 = tf.contrib.layers.batch_norm(relu2_2)
# 56,56
pool2 = tf.nn.max_pool(hyper2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
# 28,28
conv3_1 = tf.nn.conv2d(pool2, Model.weights['conv3_1_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv3_1_b']
conv3_1 = tf.contrib.layers.batch_norm(conv3_1)
relu3_1 = tf.nn.relu(conv3_1)
conv3_2 = tf.nn.conv2d(relu3_1, Model.weights['conv3_2_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv3_2_b']
conv3_2 = tf.contrib.layers.batch_norm(conv3_2)
relu3_2 = tf.nn.relu(conv3_2)
conv3_3 = tf.nn.conv2d(relu3_2, Model.weights['conv3_3_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv3_3_b']
relu3_3 = tf.nn.relu(conv3_3)
# 28,28
hyper3 = tf.contrib.layers.batch_norm(relu3_3)
# 28,28
pool3 = tf.nn.max_pool(hyper3, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
# 14,14
conv4_1 = tf.nn.conv2d(pool3, Model.weights['conv4_1_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv4_1_b']
conv4_1 = tf.contrib.layers.batch_norm(conv4_1)
relu4_1 = tf.nn.relu(conv4_1)
conv4_2 = tf.nn.conv2d(relu4_1, Model.weights['conv4_2_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv4_2_b']
conv4_2 = tf.contrib.layers.batch_norm(conv4_2)
relu4_2 = tf.nn.relu(conv4_2)
conv4_3 = tf.nn.conv2d(relu4_2, Model.weights['conv4_3_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['conv4_3_b']
relu4_3 = tf.nn.relu(conv4_3)
# 14,14
# pool4 = tf.nn.max_pool(relu4_3, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
hyper4 = tf.contrib.layers.batch_norm(relu4_3)
# 14,14
temp_conv1 = tf.nn.conv2d(hyper4, Model.weights['temp_conv1_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['temp_conv1_b']
# 14,14
deconv1 = tf.layers.conv2d_transpose(temp_conv1, 192, [3,3], strides=(2,2), padding='same') # referred from: http://warmspringwinds.github.io/tensorflow/tf-slim/2016/11/22/upsampling-and-image-segmentation-with-tensorflow-and-tf-slim/
add1 = tf.add(hyper3, deconv1)
temp_conv2 = tf.nn.conv2d(add1, Model.weights['temp_conv2_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['temp_conv2_b']
deconv2 = tf.layers.conv2d_transpose(temp_conv2, 128, [3,3], strides=(2,2), padding='same') # referred from: http://warmspringwinds.github.io/tensorflow/tf-slim/2016/11/22/upsampling-and-image-segmentation-with-tensorflow-and-tf-slim/
add2 = tf.add(hyper2, deconv2)
temp_conv3 = tf.nn.conv2d(add2, Model.weights['temp_conv3_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['temp_conv3_b']
deconv3 = tf.layers.conv2d_transpose(temp_conv3, 64, [3,3], strides=(2,2), padding='same') # referred from: http://warmspringwinds.github.io/tensorflow/tf-slim/2016/11/22/upsampling-and-image-segmentation-with-tensorflow-and-tf-slim/
add3 = tf.add(hyper1, deconv3)
temp_conv4 = tf.nn.conv2d(add3, Model.weights['temp_conv4_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['temp_conv4_b']
# deconv4 = tf.layers.conv2d_transpose(temp_conv4, 3, [3,3], strides=(2,2), padding='same') # referred from: http://warmspringwinds.github.io/tensorflow/tf-slim/2016/11/22/upsampling-and-image-segmentation-with-tensorflow-and-tf-slim/
add4 = tf.add(conv0_1, temp_conv4)
temp_conv5 = tf.nn.conv2d(add4, Model.weights['temp_conv5_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['temp_conv5_b']
uv_conv = tf.nn.conv2d(temp_conv5, Model.weights['uv_conv_W'], strides=[1,1,1,1], padding='SAME') + Model.biases['uv_conv_b']
uv_output = tf.nn.softmax(uv_conv)
return uv_output
# model = Model(32,112,112)