-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.m
169 lines (161 loc) · 6.02 KB
/
main.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
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
% implement of NMF Net
%% FCN/ACN1/ACN2
clear;clc;
net = 'acn1';% 'fcn' 'acn2'
training = 1; % 1 nmf|| 0 Inference svm || 2 svm acc|| -1 process || 4 Label info||3 layer svm
%% convexnmf/seminmf/nmf
method = 'semi'; %'convex' ||'nmf'
%% setting
batch = 2000;
if net == 'acn1'
layer = 4;
filter_size = [5,5,3,1];
stride = [2,2,2,1];
channel = [1,16,16,16,10];
elseif net == 'acn2'
layer = 6;
filter_size = [5,3,5,3,3,1];
stride = [1,2,1,2,1,1];
channel = [1,16,16,16,16,16,10];
end
r = channel;
n = filter_size;
Size_input = compute_size(28,stride,filter_size);
%% data & preprocessing
[TrainImages,TestImages,TrainLabels,TestLabels] = load_dataset('MNIST');
if training == 1
%for k = 1:floor(size(TrainImages,2)/batch)
for k = 15:-1:12
Input = cell(1,layer);
Input{1} = TrainImages(:,1+batch*(k-1):batch*(k));
W = cell(1,layer);
H = cell(1,layer);
for i = 1:layer
Input{i} = process_nmf(Input{i},filter_size(i),stride(i),channel(i),Size_input(i:i+1),batch);
fprintf('Layer %d NMF begins',i);
if method == 'semi'
[A2,Y2,~,t(i),error(i)]=seminmfrule(Input{i},r(i+1));
W{i} = A2;
H{i} = Y2;
elseif method == 'convex'
[A,Y,~,t(i),error(i)]=convexnmfrule(Input{i},r(i+1));
W{i} = TestImages*A;
H{i} = Y;
elseif method == 'nmf'
[A3,Y3,~,t(i),error(i)]=nmfrule(Input{i},r(i+1));
W{i} = A3;
H{i} = Y3;
end
%% ReLU
temp = pinv(W{i})*Input{i};
temp(temp<0) = 0;
Input{i+1} = temp;
end
filename = ['nmf_net_', num2str(k)];
save(filename,'W','H','error','t','Input');
end
elseif training == -1
%% process, combine different batch
WW = cell(1,layer);
HH = cell(1,layer);
IInput = [];
for k = 1:floor(size(TrainImages,2)/batch)
filename = ['nmf_net_', num2str(k)];
load(filename);
for i = 1:length(WW)
if k == 1
WW{i} = [WW{i} W{i}];
elseif k > 1
WW{i} = WW{i}+W{i};
end
HH{i} = [HH{i} H{i}];
end
IInput = [IInput Input{1}];
end
for i = 1:length(WW)
WW{i} = WW{i}/floor(size(TrainImages,2)/batch)
end
clear Input;
Input{1} = IInput;
clear IInput;
save('nmf_net');
elseif training == 0
%% inference
load('nmf_net');
for i = 1:layer
H_size(i) = size(H{i},2);
end
%% training set
Input{1} = single(TrainImages);
for i = 1:length(WW)
temp11 = process_nmf(reshape(Input{i},[size(Input{i},1),H_size(i),batch]),filter_size(i),stride(i),channel(i), Size_input(i:i+1),H_size(i)*batch);
Input{i} = temp11;
temp = pinv(WW{i})*Input{i};
temp(temp<0) = 0;
Input{i+1} = temp;
end
save('nmf_final_train.mat','Input','WW','HH','-v7.3')
%% testing set
IInput{1} = single(TestImages);
for i = 1:length(WW)
temp11 = process_nmf(IInput{i},filter_size(i),stride(i),channel(i), Size_input(i:i+1),H_size(i)*batch);
IInput{i} = temp11;
temp = pinv(WW{i})*IInput{i};
temp(temp<0) = 0;
IInput{i+1} = temp;
end
save('nmf_final_test.mat','IInput','WW','HH','-v7.3')
elseif training==2
load('nmf_final_train.mat')
for i = 1:layer
Input{i} = double(Input{i});
end
num = 60000;
temp_train = Input{layer+1}';
temp_train = normalize(temp_train,1,'range');% scaling
model_train = svmtrain(TrainLabels(1:num), temp_train(1:num,:),'-s 0 -t 2 -c 1 -g 0.07');
%% test transfer
load('nmf_final_test.mat')
for i = 1:layer
IInput{i} = double(IInput{i});
end
temp_test = IInput{layer+1}';
temp_test = normalize(temp_test,1,'range');
model_test = svmtrain(TestLabels, temp_test,'-s 0 -t 2 -c 1 -g 0.07');
[predict_label, accuracy, dec_values] = svmpredict(TrainLabels(1:num),temp_train(1:num,:) , model_train);
[predict_label, accuracy, dec_values] = svmpredict(TestLabels,temp_test , model_train);
[predict_label, accuracy, dec_values] = svmpredict(TrainLabels(1:num),temp_train(1:num,:) , model_test);
[predict_label, accuracy, dec_values] = svmpredict(TestLabels,temp_test , model_test);
elseif training==3
load('nmf_final_train.mat')
for i = 1:4
Input{i} = reshape(Input{i},[],60000);
model{i} = svmtrain(TrainLabels, Input{i}','-s 0 -t 2 -c 1 -g 0.07');
end
end
function Size_input = compute_size(image_size,stride,filter_size)
Size_input = [image_size, floor((image_size-(filter_size(1)-stride(1)))/stride(1))];
layer = length(stride);
for i = 2:layer
temp = floor((Size_input(end)-(filter_size(i)-stride(i)))/stride(i));
Size_input = [Size_input temp];
end
end
function Pro_image = process_nmf(input,filter_size,stride,channel, Size_input,batch)
% input is [channel Size_input^2 batch]
temp_testimage = reshape(input,[channel, Size_input(1),Size_input(1),batch]);
temp_testimage = permute(temp_testimage,[2,3,1,4]);
temp_testimage = reshape(temp_testimage,[Size_input(1),Size_input(1),channel*batch]);
%k = 1;
for i = 1:Size_input(2)
for j = 1:Size_input(2)
Pro_image((i-1)*filter_size+1:i*filter_size,(j-1)*filter_size+1:j*filter_size,:) = temp_testimage(1+(i-1)*stride:filter_size+(i-1)*stride,(j-1)*stride+1:filter_size+(j-1)*stride,:);
%Pro_image(:,:,k) = temp_testimage(1+(i-1)*stride:filter_size+(i-1)*stride,(j-1)*stride+1:filter_size+(j-1)*stride,:);
%k = k+1;
end
end
Pro_image = reshape(Pro_image,[filter_size,Size_input(2),filter_size,Size_input(2),channel,batch]);
Pro_image = permute(Pro_image,[1,3,5,2,4,6]);
Pro_image = reshape(Pro_image,[filter_size^2*channel,Size_input(2)^2*batch]);
Pro_image = [ones(1,size(Pro_image,2));Pro_image];
end