-
Notifications
You must be signed in to change notification settings - Fork 41
/
Train_Code_for_ISTA_Net_plus.py
220 lines (146 loc) · 6.9 KB
/
Train_Code_for_ISTA_Net_plus.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
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
'''
Test Platform: Tensorflow version: 1.2.0
Paper Information:
ISTA-Net: Interpretable Optimization-Inspired Deep Network for Image Compressive Sensing
Jian Zhang and Bernard Ghanem
IEEE Conference on Computer Vision and Pattern Recognition (CVPR2018), Salt Lake City, USA, Jun. 2018.
Email: [email protected]
2018/5/16
'''
import tensorflow as tf
import scipy.io as sio
import numpy as np
import os
CS_ratio = 25 # 4, 10, 25, 30,, 40, 50
if CS_ratio == 4:
n_input = 43
elif CS_ratio == 1:
n_input = 10
elif CS_ratio == 10:
n_input = 109
elif CS_ratio == 25:
n_input = 272
elif CS_ratio == 30:
n_input = 327
elif CS_ratio == 40:
n_input = 436
elif CS_ratio == 50:
n_input = 545
n_output = 1089
batch_size = 64
PhaseNumber = 5
nrtrain = 88912
learning_rate = 0.0001
EpochNum = 300
print('Load Data...')
Phi_data_Name = 'phi_0_%d_1089.mat' % CS_ratio
Phi_data = sio.loadmat(Phi_data_Name)
Phi_input = Phi_data['phi'].transpose()
Training_data_Name = 'Training_Data_Img91.mat'
Training_data = sio.loadmat(Training_data_Name)
Training_inputs = Training_data['inputs']
Training_labels = Training_data['labels']
# Computing Initialization Matrix
XX = Training_labels.transpose()
BB = np.dot(Phi_input.transpose(), XX)
BBB = np.dot(BB, BB.transpose())
CCC = np.dot(XX, BB.transpose())
PhiT_ = np.dot(CCC, np.linalg.inv(BBB))
del XX, BB, BBB, Training_data
PhiInv_input = PhiT_.transpose()
PhiTPhi_input = np.dot(Phi_input, Phi_input.transpose())
Phi = tf.constant(Phi_input, dtype=tf.float32)
PhiTPhi = tf.constant(PhiTPhi_input, dtype=tf.float32)
PhiInv = tf.constant(PhiInv_input, dtype=tf.float32)
X_input = tf.placeholder(tf.float32, [None, n_input])
X_output = tf.placeholder(tf.float32, [None, n_output])
X0 = tf.matmul(X_input, PhiInv)
PhiTb = tf.matmul(X_input, tf.transpose(Phi))
def add_con2d_weight_bias(w_shape, b_shape, order_no):
Weights = tf.get_variable(shape=w_shape, initializer=tf.contrib.layers.xavier_initializer_conv2d(), name='Weights_%d' % order_no)
biases = tf.Variable(tf.random_normal(b_shape, stddev=0.05), name='biases_%d' % order_no)
return [Weights, biases]
def ista_block(input_layers, input_data, layer_no):
tau_value = tf.Variable(0.1, dtype=tf.float32)
lambda_step = tf.Variable(0.1, dtype=tf.float32)
soft_thr = tf.Variable(0.1, dtype=tf.float32)
conv_size = 32
filter_size = 3
x1_ista = tf.add(input_layers[-1] - tf.scalar_mul(lambda_step, tf.matmul(input_layers[-1], PhiTPhi)), tf.scalar_mul(lambda_step, PhiTb)) # X_k - lambda*A^TAX
x2_ista = tf.reshape(x1_ista, shape=[-1, 33, 33, 1])
[Weights0, bias0] = add_con2d_weight_bias([filter_size, filter_size, 1, conv_size], [conv_size], 0)
[Weights1, bias1] = add_con2d_weight_bias([filter_size, filter_size, conv_size, conv_size], [conv_size], 1)
[Weights11, bias11] = add_con2d_weight_bias([filter_size, filter_size, conv_size, conv_size], [conv_size], 11)
[Weights2, bias2] = add_con2d_weight_bias([filter_size, filter_size, conv_size, conv_size], [conv_size], 2)
[Weights22, bias22] = add_con2d_weight_bias([filter_size, filter_size, conv_size, conv_size], [conv_size], 22)
[Weights3, bias3] = add_con2d_weight_bias([filter_size, filter_size, conv_size, 1], [1], 3)
x3_ista = tf.nn.conv2d(x2_ista, Weights0, strides=[1, 1, 1, 1], padding='SAME')
x4_ista = tf.nn.relu(tf.nn.conv2d(x3_ista, Weights1, strides=[1, 1, 1, 1], padding='SAME'))
x44_ista = tf.nn.conv2d(x4_ista, Weights11, strides=[1, 1, 1, 1], padding='SAME')
x5_ista = tf.multiply(tf.sign(x44_ista), tf.nn.relu(tf.abs(x44_ista) - soft_thr))
x6_ista = tf.nn.relu(tf.nn.conv2d(x5_ista, Weights2, strides=[1, 1, 1, 1], padding='SAME'))
x66_ista = tf.nn.conv2d(x6_ista, Weights22, strides=[1, 1, 1, 1], padding='SAME')
x7_ista = tf.nn.conv2d(x66_ista, Weights3, strides=[1, 1, 1, 1], padding='SAME')
x7_ista = x7_ista + x2_ista
x8_ista = tf.reshape(x7_ista, shape=[-1, 1089])
x3_ista_sym = tf.nn.relu(tf.nn.conv2d(x3_ista, Weights1, strides=[1, 1, 1, 1], padding='SAME'))
x4_ista_sym = tf.nn.conv2d(x3_ista_sym, Weights11, strides=[1, 1, 1, 1], padding='SAME')
x6_ista_sym = tf.nn.relu(tf.nn.conv2d(x4_ista_sym, Weights2, strides=[1, 1, 1, 1], padding='SAME'))
x7_ista_sym = tf.nn.conv2d(x6_ista_sym, Weights22, strides=[1, 1, 1, 1], padding='SAME')
x11_ista = x7_ista_sym - x3_ista
return [x8_ista, x11_ista]
def inference_ista(input_tensor, n, X_output, reuse):
layers = []
layers_symetric = []
layers.append(input_tensor)
for i in range(n):
with tf.variable_scope('conv_%d' %i, reuse=reuse):
[conv1, conv1_sym] = ista_block(layers, X_output, i)
layers.append(conv1)
layers_symetric.append(conv1_sym)
return [layers, layers_symetric]
[Prediction, Pre_symetric] = inference_ista(X0, PhaseNumber, X_output, reuse=False)
cost0 = tf.reduce_mean(tf.square(X0 - X_output))
def compute_cost(Prediction, X_output, PhaseNumber):
cost = tf.reduce_mean(tf.square(Prediction[-1] - X_output))
cost_sym = 0
for k in range(PhaseNumber):
cost_sym += tf.reduce_mean(tf.square(Pre_symetric[k]))
return [cost, cost_sym]
[cost, cost_sym] = compute_cost(Prediction, X_output, PhaseNumber)
cost_all = cost + 0.01*cost_sym
optm_all = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost_all)
init = tf.global_variables_initializer()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
saver = tf.train.Saver(tf.global_variables(), max_to_keep=100)
sess = tf.Session(config=config)
sess.run(init)
print("...............................")
print("Phase Number is %d, CS ratio is %d%%" % (PhaseNumber, CS_ratio))
print("...............................\n")
print("Strart Training..")
model_dir = 'Phase_%d_ratio_0_%d_ISTA_Net_plus_Model' % (PhaseNumber, CS_ratio)
output_file_name = "Log_output_%s.txt" % (model_dir)
for epoch_i in range(0, EpochNum+1):
randidx_all = np.random.permutation(nrtrain)
for batch_i in range(nrtrain // batch_size):
randidx = randidx_all[batch_i*batch_size:(batch_i+1)*batch_size]
batch_ys = Training_labels[randidx, :]
batch_xs = np.dot(batch_ys, Phi_input)
feed_dict = {X_input: batch_xs, X_output: batch_ys}
sess.run(optm_all, feed_dict=feed_dict)
output_data = "[%02d/%02d] cost: %.4f, cost_sym: %.4f \n" % (epoch_i, EpochNum, sess.run(cost, feed_dict=feed_dict), sess.run(cost_sym, feed_dict=feed_dict))
print(output_data)
output_file = open(output_file_name, 'a')
output_file.write(output_data)
output_file.close()
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if epoch_i <= 30:
saver.save(sess, './%s/CS_Saved_Model_%d.cpkt' % (model_dir, epoch_i), write_meta_graph=False)
else:
if epoch_i % 20 == 0:
saver.save(sess, './%s/CS_Saved_Model_%d.cpkt' % (model_dir, epoch_i), write_meta_graph=False)
print("Training Finished")
sess.close()