-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathDDZO_DGT.m
69 lines (48 loc) · 2.08 KB
/
DDZO_DGT.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
function [xminuxbar, sq_grad,time] = DDZO_DGT(W_aug, x, n,N,gc,cf, lambda,aalpha, features, labels,bs, iter_num)
fprintf('Starting DDZO_DGT\n');
sq_grad = zeros(iter_num, 1);
xminuxbar = zeros(iter_num, 1);
time = zeros(iter_num, 1);
gs = zeros(n*N, 1);
ss = zeros(n*N, 1);
eta1 = 1.3;
u1 = 0.1;
upd = textprogressbar(iter_num);
for t = 2:iter_num
tic;
upd(t);
eta = eta1;
u = u1 / t^(3/4);
prev_gs = gs;
gradient = zeros(N*n,1);
for ii = 1 : N
std_basis = eye(n);
grad_temp = zeros(n, 1);
for zz = 1:n
for jj=(ii-1)*bs+1:ii*bs
grad_temp = grad_temp + (cf(x((ii-1)*n+1:ii*n,t-1)+u*std_basis(:, zz),lambda,aalpha, features(:,jj), labels(jj),bs, N) - cf(x((ii-1)*n+1:ii*n,t-1)-u*std_basis(:, zz),lambda,aalpha, features(:,jj), labels(jj),bs, N)).*std_basis(:, zz)/u/2;
end
end
gradient((ii-1)*n+1:ii*n) = gradient((ii-1)*n+1:ii*n) + grad_temp;
end
gs = gradient;
ss = W_aug* ( ss + gs - prev_gs ) ;
x(:,t) = W_aug*( x(:, t -1) - eta*ss );
xs = reshape(x(:,t), [n, N]);
x_avg = sum(xs, 2) / N;
for k = 1:N
xminuxbar(t) = xminuxbar(t)+(norm(xs(:, k)-x_avg))^2;
% fminufstar(t) = fminufstar(t)+loss_func(x_avg, y{k}, a_Re{k}, a_Im{k});
end
temp_grad = zeros(N*n,1);
for ii = 1 : N
for jj=(ii-1)*bs+1:ii*bs
temp_grad((ii-1)*n+1:ii*n) = temp_grad((ii-1)*n+1:ii*n) + gc(x_avg,lambda,aalpha, features(:,jj), labels(jj),bs, N); % This is compute the gradient in each node, batch_size works here.
end
end
g = reshape(temp_grad, [n, N]);
sq_grad(t) = sum(sum(g, 2).^2);
t_temp = toc;
time(t) = time(t - 1) + t_temp;
end
end