-
Notifications
You must be signed in to change notification settings - Fork 0
/
contextNetwork_2.m
133 lines (110 loc) · 4.57 KB
/
contextNetwork_2.m
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
function [net, layer_out, classifier_outs] = contextNetwork_2(net, layer_in, ...
ker_size, nh0, nh, nClass, layers, newLr, layer_prefix, recursive,dilate)
classifier_outs = [];
if layers == 0
layer_out = layer_in;
return;
end
layer_out=cell(1,layers);
% shared version for recursive net
if recursive
cw_param_f = sprintf('%s_cw_f_shared', layer_prefix);
cw_param_b = sprintf('%s_cw_b_shared', layer_prefix);
cw_f_shared = 1e-2*randn(ker_size, ker_size, nh0, nh, 'single');
cw_b_shared = zeros(1, 1, nh, 'single');
classifier_f = sprintf('%s_cw_classifier_f_shared', layer_prefix);
classifier_b = sprintf('%s_cw_classifier_b_shared', layer_prefix);
classifier_f_shared = 1e-2*randn(1, 1, nh, nClass, 'single');
classifier_b_shared = zeros(1, 1, nClass, 'single');
end
for i = 1 : layers
if i == 1,
conv_in = layer_in;
end
conv_layer = sprintf('%s_cw_conv_%d', layer_prefix, i);
relu_layer = sprintf('%s_cw_relu_%d',layer_prefix, i);
drop_layer = sprintf('%s_cw_drop_%d',layer_prefix, i);
conv_out = sprintf('%s_conv_out_%d', layer_prefix, i);
relu_out = sprintf('%s_relu_out_%d', layer_prefix, i);
drop_out = sprintf('%s_drop_out_%d', layer_prefix, i);
bn_layer = sprintf('%s_cw_bn_%d', layer_prefix, i);
bn_out = sprintf('%s_bn_out_%d', layer_prefix, i);
if ~recursive
cw_param_f = sprintf('%s_cw_f_%d', layer_prefix, i);
cw_param_b = sprintf('%s_cw_b_%d', layer_prefix, i);
cw_f_shared = 1e-2*randn(ker_size, ker_size, nh0, nh, 'single');
cw_b_shared = zeros(1, 1, nh, 'single');
end
%% conv layer
net.addLayer(conv_layer, ...
dagnn.Conv('size', [ker_size ker_size nh0 nh], 'pad', floor((ker_size+(dilate(i)-1)*(ker_size-1))/2),'dilate',dilate(i)), ...
conv_in, conv_out, {cw_param_f,cw_param_b});
f = net.getParamIndex(cw_param_f) ;
net.params(f).value = cw_f_shared ;
net.params(f).learningRate = 1 * newLr;
net.params(f).weightDecay = 1 ;
f = net.getParamIndex(cw_param_b) ;
net.params(f).value = cw_b_shared ;
net.params(f).learningRate = 2 * newLr ;
net.params(f).weightDecay = 1 ;
%% Batch Normalization
bn_param_f = sprintf('%s_bn_f_%d', layer_prefix, i);
bn_param_b = sprintf('%s_bn_b_%d', layer_prefix, i);
bn_param_m = sprintf('%s_bn_m_%d', layer_prefix, i);
net.addLayer(bn_layer, ...
dagnn.BatchNorm(), ...
conv_out, bn_out, {bn_param_f, bn_param_b, bn_param_m});
f = net.getParamIndex(bn_param_f) ;
net.params(f).value = ones(nh, 1, 'single') ;
net.params(f).learningRate = 1 * newLr;
net.params(f).weightDecay = 1 ;
f = net.getParamIndex(bn_param_b) ;
net.params(f).value = zeros(nh, 1, 'single') ;
net.params(f).learningRate = 1 * newLr ;
net.params(f).weightDecay = 1 ;
f = net.getParamIndex(bn_param_m) ;
net.params(f).value = zeros(nh, 2, 'single') ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 0 ;
%% ReLU
net.addLayer(relu_layer, ...
dagnn.ReLU(),...
bn_out, relu_out);
conv_in = relu_out; % input for next conv layer
nh0 = nh;
%% dropout
% net.addLayer(drop_layer, ...
% dagnn.DropOut( 'rate', i/layers * 0.5),...
% relu_out, drop_out);
%
%% add an output layer
% [net, classifier_out] = skipNetwork(net, {relu_out}, nh, nh, ...
% nClass, newLr, sprintf('skip5_%d',i));
% classifier = sprintf('%s_cw_classifier_%d', layer_prefix, i);
% classifier_out = sprintf('%s_cw_classifier_out_%d', layer_prefix, i);
%
% if ~recursive
% classifier_f = sprintf('%s_cw_classifier_f_%d', layer_prefix, i);
% classifier_b = sprintf('%s_cw_classifier_b_%d', layer_prefix, i);
% classifier_f_shared = 1e-2*randn(1, 1, nh, nClass, 'single');
% classifier_b_shared = zeros(1, 1, nClass, 'single');
% end
%
%
% net.addLayer(classifier, ...
% dagnn.Conv('size', [1 1 nh nClass], 'pad', 0), ...
% relu_out, classifier_out, {classifier_f,classifier_b});
%
% f = net.getParamIndex(classifier_f) ;
% net.params(f).value = classifier_f_shared;
% net.params(f).learningRate = 1 * newLr ;
% net.params(f).weightDecay = 1 ;
%
% f = net.getParamIndex(classifier_b) ;
% net.params(f).value = classifier_b_shared;
% net.params(f).learningRate = 2 * newLr;
% net.params(f).weightDecay = 1 ;
% classifier_out = mat2cell(classifier_out, 1);
layer_out{i} = relu_out;
end
% layer_out = relu_out;