-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathCSRNet.py
More file actions
153 lines (128 loc) · 6.88 KB
/
CSRNet.py
File metadata and controls
153 lines (128 loc) · 6.88 KB
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
import numpy as np
import paddle
import paddle.fluid as fluid
from paddle.fluid.layer_helper import LayerHelper
from paddle.fluid.dygraph.nn import Conv2D, Pool2D, BatchNorm, FC
from paddle.fluid.dygraph.base import to_variable
import pickle
import pickle
with open('/home/aistudio/work/vgg16.pkl', 'rb') as f:
mode = pickle.load(f)
#print(mode.keys())
# 【0， 2， 5， 7， 10， 12， 14， 17， 19， 21】
# features.0.weight
class VGG16(fluid.dygraph.Layer):
def __init__(self, name_scope, cfg, batch_norm=False, dilation=False):
super(VGG16, self).__init__(name_scope)
self.cfg = cfg
self.batch_norm = batch_norm
self.dilation = dilation
self.layers = self.make_layers(self.cfg)
#print(self.layers[0])
def forward(self, x):
for op in self.layers:
x = op(x)
return x
def make_layers(self, cfg, batch_norm=False, dilation=False):
'''
cfg: 参数
batch_norm: 是否正则化
dilation: 膨化系数， frontend:1, backend:2
'''
if dilation:
d_rate = 2
else:
d_rate = 1
layers = []
i = 0
for v in cfg:
if v == 'M':
layers.append(self.add_sublayer('pool_' + str(i), Pool2D(self.full_name(), pool_size=2, pool_stride=2)))
else:
#layers.append(Conv2D(num_filters=v, filter_size=3, padding=d_rate, dilation=d_rate))
if batch_norm:
layers.append(self.add_sublayer('conv_' + str(i), Conv2D(self.full_name(), num_filters=v, filter_size=3, padding=d_rate, dilation=d_rate,param_attr=fluid.ParamAttr(
initializer=fluid.initializer.NormalInitializer(scale=0.01) ), bias_attr=fluid.ParamAttr())))
layers.append(self.add_sublayer('batch_norm_' + str(i), BatchNorm(v, act="relu",param_attr=fluid.ParamAttr(
initializer=fluid.initializer.ConstantInitializer(value=1) ), bias_attr=fluid.ParamAttr(
initializer=fluid.initializer.ConstantInitializer(value=0)))))
else:
layers.append(self.add_sublayer('conv_' + str(i), Conv2D(self.full_name(), num_filters=v, filter_size=3, padding=d_rate, dilation=d_rate, act='relu', param_attr=fluid.ParamAttr(
initializer=fluid.initializer.NormalInitializer(scale=0.01) ), bias_attr=fluid.ParamAttr())))
i+=1
return layers
class FrontEnd(fluid.dygraph.Layer):
def __init__(self, name_scope, cfg, batch_norm=False, dilation=False):
super(FrontEnd, self).__init__(name_scope)
# 【0， 2， 5， 7， 10， 12， 14， 17， 19， 21】
self.name = ['features.0', 'features.2', 'M', 'features.5', 'features.7', 'M', 'features.10', 'features.12', 'features.14', 'M', 'features.17', 'features.19', 'features.21']
self.layers = self.make_layers(cfg, self.name)
def forward(self, x):
for op in self.layers:
x = op(x)
#print(op.bias)
return x
def make_layers(self, cfg, name, batch_norm=False, dilation=False):
'''
cfg: 参数
batch_norm: 是否正则化
dilation: 膨化系数， frontend:1, backend:2
'''
if dilation:
d_rate = 2
else:
d_rate = 1
layers = []
for v, n in zip(cfg, name):
if v == 'M':
layers.append(self.add_sublayer('pool_' + str(n), Pool2D(self.full_name(), pool_size=2, pool_stride=2)))
else:
#layers.append(Conv2D(num_filters=v, filter_size=3, padding=d_rate, dilation=d_rate))
if batch_norm:
layers.append(self.add_sublayer('conv_' + str(n), Conv2D(self.full_name(), num_filters=v, filter_size=3, padding=d_rate, dilation=d_rate,param_attr=fluid.ParamAttr(
initializer=fluid.initializer.NumpyArrayInitializer(mode[n + '.weight'])), bias_attr=fluid.ParamAttr(
initializer=fluid.initializer.NumpyArrayInitializer(mode[n + '.bias'])) )))
layers.append(self.add_sublayer('batchnorm_' + str(n), BatchNorm(v, act="relu",param_attr=fluid.ParamAttr(
initializer=fluid.initializer.ConstantInitializer(value=1) ), bias_attr=fluid.ParamAttr(
initializer=fluid.initializer.ConstantInitializer(value=0)))))
else:
layers.append(self.add_sublayer('conv_' + str(n), Conv2D(self.full_name(), num_filters=v, filter_size=3, padding=d_rate, dilation=d_rate, act='relu', param_attr=fluid.ParamAttr(
initializer=fluid.initializer.NumpyArrayInitializer(mode[n + '.weight'])),bias_attr=fluid.ParamAttr(
initializer=fluid.initializer.NumpyArrayInitializer(mode[n + '.bias'])) )))
return layers
class CSRNet(fluid.dygraph.Layer):
def __init__(self, name_scope):
super(CSRNet, self).__init__(name_scope)
# 'M' 池化
self.frontend_feat=[64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512]
self.backend_feat=[512, 512, 512, 256, 128, 64]
self.frontend = self.add_sublayer('VGG16', FrontEnd('frontend', self.frontend_feat))
self.backend = self.add_sublayer('backend', VGG16('backend', self.backend_feat, batch_norm=True, dilation=True))
self.output_layer = Conv2D(self.full_name(), num_filters=1, filter_size=1, param_attr=fluid.ParamAttr(
initializer=fluid.initializer.NormalInitializer(scale=0.01) ), bias_attr=fluid.ParamAttr(
initializer=fluid.initializer.ConstantInitializer(value=0)))
def forward(self, x):
x = self.frontend(x) # NCHW
x = self.backend(x)
x = self.output_layer(x)
return x
'''
'''
if __name__ == '__main__':
with fluid.dygraph.guard():
net = CSRNet('csrnet')
img = np.random.rand(1, 3, 256, 512).astype('float32')
label = np.zeros([1, 1, 64, 128]).astype('float32')
img = fluid.dygraph.to_variable(img)
label = fluid.dygraph.to_variable(label)
outs = net(img)
print(net.state_dict().keys())
print(len(list(net.state_dict().keys())))
#outs.backward()
#print(net.state_dict().keys())
#loss = mse_loss(outs, label)
#mean_loss = fluid.layers.mean(loss)
#loss.backward()
#print(outs.numpy())
#print(loss)
#print("acc", evaluate(outs, label))