-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_model.m
179 lines (157 loc) · 8.29 KB
/
train_model.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
170
171
172
173
174
175
176
177
178
179
v_indices = 1:visible;
h_indices = visible + 1 : visible + hidden;
vh_indices = [v_indices, h_indices];
% training_indices = find((partition == 1) .* ((label == label_train(1)) + (label == label_train(2))));
% validation_indices = find((partition == 2) .* ((label == label_train(1)) + (label == label_train(2))));
% cheat_indices = find((label ~= label_train(1)) .* (label ~= label_train(2)));
training_indices = find((partition == 1) .* (label == label_train));
validation_indices = find((partition == 2) .* (label == label_train));
cheat_indices = find(label ~= label_train);
cheat_indices = datasample(cheat_indices, 1000);
positive_vhbs = double(randn(visible + hidden + 1, const_samples) < 0.1);
positive_vhbs(visible + hidden + 1, :) = 1;
negative_vhbs = double(randn(visible + hidden + 1, const_samples) < 0.1);
negative_vhbs(visible + hidden + 1, :) = 1;
persistent_samples = randn(visible + hidden, const_samples);
gradient_history = zeros(1, const_iteration / print_iteration);
training_free_energy_history = zeros(1, const_iteration / print_iteration);
validation_free_energy_history = zeros(1, const_iteration / print_iteration);
cheat_free_energy_history = zeros(1, const_iteration / print_iteration);
history_i = 1;
training_data = gpuArray(single(data(training_indices, :)));
validation_data = gpuArray(single(data(validation_indices, :)));
cheat_data = gpuArray(single(data(cheat_indices, :)));
% training_data = gpuArray(single(zeros(length(training_indices), visible)));
% validation_data = gpuArray(single(zeros(length(validation_indices), visible)));
% cheat_data = gpuArray(single(zeros(length(cheat_indices), visible)));
%% prepare training data
% for index = 1:length(training_indices)
% train_index = training_indices(index);
% training_data(index, :) = get_data_from_index(data_all, train_index, const_h, const_w, channels)';
% end
% for index = 1:length(validation_indices)
% val_index = validation_indices(index);
% validation_data(index, :) = get_data_from_index(data_all, val_index, const_h, const_w, channels)';
% end
% for index = 1:length(cheat_indices)
% cheat_index = cheat_indices(index);
% cheat_data(index, :) = get_data_from_index(data_all, cheat_index, const_h, const_w, channels)';
% end
for iteration = 1:const_iteration
%% prepare mini-batch data
vn = datasample(training_data, const_samples, 'Replace', true)';
if method == 1
%% positive gibbs
positive_vhbs(1:visible, 1:const_samples) = vn;
for gibbs = 1:const_gibbs_positive
index = h_indices(randi(numel(h_indices)));
result = matrix(index, :) * positive_vhbs;
prob = 1 ./ (1 + exp(-result));
positive_vhbs(index,:) = rand() < prob;
end
dmatrix_positive = positive_vhbs * positive_vhbs' / const_samples;
%% negative steps
for gibbs = 1:const_gibbs_negative
index = vh_indices(randi(numel(vh_indices)));
result = matrix(index, :) * negative_vhbs;
prob = 1 ./ (1 + exp(-result));
negative_vhbs(index,:) = rand() < prob;
end
dmatrix_negative = negative_vhbs * negative_vhbs' / const_samples;
dmatrix = dmatrix_positive - dmatrix_negative;
elseif method == 2
%% contrastive divergence mean field
mhn = gpuArray(single(ones(hidden, const_samples) * 0.5));
for i = 1:const_cd_0% * hidden
% index = randi(hidden);
index = 1:hidden;
temp = matrix(visible + index, :) * [vn; mhn; ones(1, const_samples)];
mhn(index, :) = (1 - momentum) * mhn(index, :) + momentum * 1 ./ (1 + exp(-temp));
end
mvn = vn;
for i = 1:const_cd_1% * visible
% index = randi(visible);
index = 1:visible;
temp = matrix(index, :) * [mvn; mhn; ones(1, const_samples)];
mvn(index, :) = (1 - momentum) * mvn(index, :) + momentum * 1 ./ (1 + exp(-temp));
end
mhnt = mhn;
for i = 1:const_cd_2% * hidden
% index = randi(hidden);
index = 1:hidden;
temp = matrix(visible + index, :) * [mvn; mhnt; ones(1, const_samples)];
mhnt(index, :) = (1 - momentum) * mhnt(index, :) + momentum * 1 ./ (1 + exp(-temp));
end
dW = (mhn * mhn' - mhnt * mhnt') ./ const_samples - weight_decay .* matrix(visible+1:visible+hidden, visible+1:visible+hidden);
dV = (vn * vn' - mvn * mvn') ./ const_samples - weight_decay .* matrix(1:visible, 1:visible);
dJ = (vn * mhnt' - mvn * mhnt') ./ const_samples - weight_decay .* matrix(1:visible, visible+1:visible+hidden);
dBv = mean(vn - mvn, 2); % no wd for bias
dBh = mean(mhn - mhnt, 2); % no wd for bias
dmatrix = [dV, dJ, dBv; dJ', dW, dBh; dBv', dBh', 0];
elseif method == 3
end
%% updates
temp_lr = 0.1 ^ sum((iteration / const_iteration) > lr_decay) * learning_rate;
dmatrix = dmatrix .* mask;
matrix = matrix + dmatrix .* temp_lr;
%% print
if mod(iteration, print_iteration) == 0
% visualize_model(matrix, visible, hidden, const_h, fig1, fig2, fig3, disp_scale);
matrix_sum = gather(sum(sum(abs(dmatrix))));
matrix_max = gather(max(max(abs(dmatrix))));
disp([datestr(now), ' iteration ', num2str(iteration), ', learning_rate ', num2str(temp_lr), ', gradient_sum ', num2str(matrix_sum), ', gradient_max ', num2str(matrix_max)]);
gradient_history(history_i) = log10(max(matrix_sum, 0.0000001));
% training_free_energy_history(history_i) = training_free_energy;
% validation_free_energy_history(history_i) = validation_free_energy;
% cheat_free_energy_history(history_i) = cheat_free_energy;
history_i = history_i + 1;
% figure(fig_history);
% plot([training_free_energy_history;validation_free_energy_history;cheat_free_energy_history]');
% legend('training', 'validation', 'cheat'); drawnow;
% plot(gradient_history); drawnow;
end
%% all visible case free energy over all training data
if mod(iteration, energy_iteration) == 0 && (type == 1 || type == 2)
training_free_energies = 99999999 * ones(length(training_indices), 1);
for index = 1:length(training_indices)
visible_binary = training_data(index, :);
for hidden_int = 0:2^hidden-1
hidden_binary = de2bi(hidden_int, hidden, 'left-msb');
temp_data = [visible_binary, hidden_binary, 1];
training_free_energies(index) = min(training_free_energies(index), gather(-temp_data * matrix * temp_data') / 2);
end
end
training_free_energy = mean(training_free_energies / 2^hidden);
validation_free_energies = 99999999 * ones(length(validation_indices), 1);
for index = 1:length(validation_indices)
visible_binary = validation_data(index, :);
for hidden_int = 0:2^hidden-1
hidden_binary = de2bi(hidden_int, hidden, 'left-msb');
temp_data = [visible_binary, hidden_binary, 1];
validation_free_energies(index) = min(validation_free_energies(index), gather(-temp_data * matrix * temp_data') / 2);
end
end
validation_free_energy = mean(validation_free_energies / 2^hidden);
cheat_free_energies = 99999999 * ones(length(cheat_indices), 1);
for index = 1:length(cheat_indices)
visible_binary = cheat_data(index, :);
for hidden_int = 0:2^hidden-1
hidden_binary = de2bi(hidden_int, hidden, 'left-msb');
temp_data = [visible_binary, hidden_binary, 1];
cheat_free_energies(index) = min(cheat_free_energies(index), gather(-temp_data * matrix * temp_data') / 2);
end
end
cheat_free_energy = mean(cheat_free_energies / 2^hidden);
disp(['training ', num2str(training_free_energy), ', validation ', num2str(validation_free_energy), ', cheat ', num2str(cheat_free_energy)]);
end
%% save model
if mod(iteration, store_iteration) == 0
save(['matrix_', num2str(label_train), '_lr_', num2str(learning_rate),'_mom_', num2str(momentum), '_hidden_',num2str(hidden),'.mat'],'matrix');
end
%% break point
if iteration == 1000000
a = 0;
end
end
% colorbar;
% imshow(kron(matrix,ones(10)), [-50, 50], 'Colormap', jet(256));