lenet-5训练MNIST

function checkgrads(net,x,y) epsilon = 1e-4; er = 1e-8; n = numel(net.layers); for l = n : -1 : 2 if strcmp(net.layers{l}.type,'Soft') for j = 1 : numel(net.layers{l}.b) net_m = net; net_p = net; net_p.layers{l}.b(j) = net_m.layers{l}.b(j) + epsilon; net_m.layers{l}.b(j) = net_m.layers{l}.b(j) - epsilon; net_m = cnnff(net_m,x);net_m = cnnbp(net_m,y); net_p = cnnff(net_p,x);net_p = cnnbp(net_p,y); d = (net_p.L - net_m.L) / (2 * epsilon); e = abs(d - net.layers{l}.db(j)); if e > er error('numrical graddient checking failed at softb'); end end for i = 1 : size(net.layers{l}.w,1) for j = 1 : size(net.layers{l}.w,2) net_m = net; net_p = net; net_p.layers{l}.w(i,j) = net_p.layers{l}.w(i,j) + epsilon; net_m.layers{l}.w(i,j) = net_m.layers{l}.w(i,j) - epsilon; net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); d = (net_p.L - net_m.L)/(2*epsilon); e = abs(d - net.layers{l}.dw(i,j)); if e > er error('numrical graddient checking faild at softw'); end end end end if strcmp(net.layers{l}.type,'F6') for j = 1 : size(net.layers{l}.b,1) net_m = net;net_p = net; net_p.layers{l}.b(j) = net_p.layers{l}.b(j) + epsilon; net_m.layers{l}.b(j) = net_m.layers{l}.b(j) - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); %db = net.layers{l}.db(j); d = (net_p.L - net_m.L) / (2 * epsilon); e = abs(d - net.layers{l}.db(j)); if e > er error('numrical graddient checking failed at F6b'); end end for i = 1 : size(net.layers{l}.w,1) for j = 1 : size(net.layers{l}.w,2) net_m = net;net_p = net; net_p.layers{l}.w(i,j) = net_p.layers{l}.w(i,j) + epsilon; net_m.layers{l}.w(i,j) = net_m.layers{l}.w(i,j) - epsilon; net_p = cnnff(net_p,x);net_p =cnnbp(net_p,y); net_m = cnnff(net_m,x);net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2 * epsilon); e = abs(d - net.layers{l}.dw(i,j)); if e > er error('numrical gradient checking failed at F6w'); end end end end if strcmp(net.layers{l}.type,'C5') for i = 1 : size(net.layers{l}.w,1) for j = 1 : size(net.layers{l}.w,2) net_m = net;net_p = net; net_p.layers{l}.w(i,j) = net_p.layers{l}.w(i,j) + epsilon; net_m.layers{l}.w(i,j) = net_m.layers{l}.w(i,j) - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2 * epsilon); e = abs(d - net.layers{l}.dw(i,j)); if e > er error('numrical gradient checking failed at C5w'); end end end for j = 1 : size(net.layers{l}.b,1) net_m = net;net_p = net; net_p.layers{l}.b(j) = net_p.layers{l}.b(j) + epsilon; net_m.layers{l}.b(j) = net_m.layers{l}.b(j) - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2 * epsilon); e = abs(d - net.layers{l}.b(j)); if e > er error('numrical gradient checking failed at C5b'); end end end if strcmp(net.layers{l}.type,'C3') for j = 1 : numel(net.layers{l}.a) net_m = net; net_p = net; net_p.layers{l}.b{j} = net_p.layers{l}.b{j} + epsilon; net_m.layers{l}.b{j} = net_m.layers{l}.b{j} - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2*epsilon); e = abs(d - net.layers{l}.db{j}); if e > er error('numrical gradient checking failed at C3b'); end for i = 1 : numel(net.layers{l-1}.a) for u = 1 : size(net.layers{l}.w{i}{j},1) for v = 1 : size(net.layers{l}.w{i}{j},2) net_m = net; net_p = net; net_p.layers{l}.w{i}{j}(u,v) = net_p.layers{l}.w{i}{j}(u,v) + epsilon; net_m.layers{l}.w{i}{j}(u,v) = net_m.layers{l}.w{i}{j}(u,v) - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2*epsilon); e = abs(d - net.layers{l}.dw{i}{j}(u,v)); if e > er error('numrical gradient checking failed at C3w'); end end end end end end if strcmp(net.layers{l}.type,'C1') for j = 1 : numel(net.layers{l}.a) net_m = net; net_p = net; net_p.layers{l}.b{j} = net_p.layers{l}.b{j} + epsilon; net_m.layers{l}.b{j} = net_m.layers{l}.b{j} - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2*epsilon); e = abs(d - net.layers{l}.db{j}); if e > er error('numrical gradient checking failed at C1b'); end for i = 1 : numel(net.layers{l-1}.a) for u = 1 : size(net.layers{l}.w{i}{j},1) for v = 1 : size(net.layers{l}.w{i}{j},2) net_m = net; net_p = net; net_p.layers{l}.w{i}{j}(u,v) = net_p.layers{l}.w{i}{j}(u,v) + epsilon; net_m.layers{l}.w{i}{j}(u,v) = net_m.layers{l}.w{i}{j}(u,v) - epsilon; net_p = cnnff(net_p,x); net_p = cnnbp(net_p,y); net_m = cnnff(net_m,x); net_m = cnnbp(net_m,y); d = (net_p.L - net_m.L) / (2*epsilon); e = abs(d - net.layers{l}.dw{i}{j}(u,v)); if e > er error('numrical gradient checking failed at C1w'); end end end end end end end end