-
Notifications
You must be signed in to change notification settings - Fork 53
/
generate_init_params.py
97 lines (82 loc) · 3.44 KB
/
generate_init_params.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
import tensorflow as tf
import sys
import numpy as np
sys.path.append("/home/xiaobai/caffe/python")
import caffe
from vggm import vggM
lr = tf.placeholder(tf.float32, shape=())
net_caffe = caffe.Net("./VGGM.prototxt", "./VGG_CNN_M.caffemodel", caffe.TEST)
caffe_layers = {}
for i, layer in enumerate(net_caffe.layers):
layer_name = net_caffe._layer_names[i]
caffe_layers[layer_name] = layer
def caffe_weights(layer_name):
layer = caffe_layers[layer_name]
return layer.blobs[0].data
def caffe_bias(layer_name):
layer = caffe_layers[layer_name]
return layer.blobs[1].data
conv1_w = caffe_layers['conv1'].blobs[0].data.transpose((2, 3, 1, 0))
conv2_w = caffe_layers['conv2'].blobs[0].data.transpose((2, 3, 1, 0))
conv3_w = caffe_layers['conv3'].blobs[0].data.transpose((2, 3, 1, 0))
conv4_w = caffe_layers['conv4'].blobs[0].data.transpose((2, 3, 1, 0))
conv5_w = caffe_layers['conv5'].blobs[0].data.transpose((2, 3, 1, 0))
fc6_w = caffe_layers['fc6'].blobs[0].data.transpose((1,0))
fc7_w = caffe_layers['fc7'].blobs[0].data.transpose((1,0))
conv1_b = caffe_layers['conv1'].blobs[1].data
conv2_b = caffe_layers['conv2'].blobs[1].data
conv3_b = caffe_layers['conv3'].blobs[1].data
conv4_b = caffe_layers['conv4'].blobs[1].data
conv5_b = caffe_layers['conv5'].blobs[1].data
fc6_b = caffe_layers['fc6'].blobs[1].data
fc7_b = caffe_layers['fc7'].blobs[1].data
parameters_dict = {}
parameters_dict['conv1_w'] = conv1_w
parameters_dict['conv2_w'] = conv2_w
parameters_dict['conv3_w'] = conv3_w
parameters_dict['conv4_w'] = conv4_w
parameters_dict['conv5_w'] = conv5_w
parameters_dict['fc6_w'] = fc6_w
parameters_dict['fc7_w'] = fc7_w
parameters_dict['conv1_b'] = conv1_b
parameters_dict['conv2_b'] = conv2_b
parameters_dict['conv3_b'] = conv3_b
parameters_dict['conv4_b'] = conv4_b
parameters_dict['conv5_b'] = conv5_b
parameters_dict['fc6_b'] = fc6_b
parameters_dict['fc7_b'] = fc7_b
model = vggM()
inputs = tf.placeholder(dtype=tf.float32,shape=(128,107,107,3))
label = tf.placeholder(dtype=tf.uint8,shape=(128,2))
outputs = model.vggM(inputs)
lossOp = model.loss(outputs,label)
optimizer = tf.train.MomentumOptimizer(learning_rate=0.0002,momentum=0.9)
saver = tf.train.Saver(max_to_keep=40)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
with tf.variable_scope('VGGM'):
with tf.variable_scope("layer1") as scope:
scope.reuse_variables()
weights = tf.get_variable("weights")
biases = tf.get_variable("biases")
sess.run(weights.assign(parameters_dict["conv1_w"]))
sess.run(biases.assign(parameters_dict["conv1_b"]))
with tf.variable_scope("layer2") as scope:
scope.reuse_variables()
weights = tf.get_variable("weights")
biases = tf.get_variable("biases")
sess.run(weights.assign(parameters_dict["conv2_w"]))
sess.run(biases.assign(parameters_dict["conv2_b"]))
with tf.variable_scope("layer3") as scope:
scope.reuse_variables()
weights = tf.get_variable("weights")
biases = tf.get_variable("biases")
sess.run(weights.assign(parameters_dict["conv3_w"]))
sess.run(biases.assign(parameters_dict["conv3_b"]))
with tf.variable_scope("layer4") as scope:
scope.reuse_variables()
weights = tf.get_variable("weights")
biases = tf.get_variable("biases")
sess.run(weights.assign(parameters_dict["conv4_w"]))
sess.run(biases.assign(parameters_dict["conv4_b"]))
saveRes = saver.save(sess, 'vggMParams.ckpt')