Mercurial > hg > camir-aes2014

annotate toolboxes/RBM/training_rbm.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 function [W visB hidB] = training_rbm(conf,W,data_file)
wolffd@0 2 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 3 % Training RBM %
wolffd@0 4 % conf: training setting %
wolffd@0 5 % W: weights of connections %
wolffd@0 6 % -*-sontran2012-*- %
wolffd@0 7 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
wolffd@0 8 %% load data
wolffd@0 9 vars = whos('-file', data_file);
wolffd@0 10 A = load(data_file,vars(1).name);
wolffd@0 11 data = A.(vars(1).name);
wolffd@0 12 assert(~isempty(data),'[KBRBM] Data is empty');
wolffd@0 13 %% initialization
wolffd@0 14 visNum = size(data,2);
wolffd@0 15 hidNum = conf.hidNum;
wolffd@0 16 sNum = conf.sNum;
wolffd@0 17 lr = conf.params(1);
wolffd@0 18 N = 10; % Number of epoch training with lr_1
wolffd@0 19 W = 0.1*randn(visNum - size(W,1),size(W,2));
wolffd@0 20 W = 0.1*randn(size(W,1),hidNum-size(W,2));
wolffd@0 21
wolffd@0 22 DW = zeros(size(W));
wolffd@0 23 visB = zeros(1,visNum);
wolffd@0 24 DVB = zeros(1,visNum);
wolffd@0 25 hidB = zeros(1,hidNum);
wolffd@0 26 DHB = zeros(1,hidNum);
wolffd@0 27 visP = zeros(sNum,visNum);
wolffd@0 28 visN = zeros(sNum,visNum);
wolffd@0 29 visNs = zeros(sNum,visNum);
wolffd@0 30 hidP = zeros(sNum,hidNum);
wolffd@0 31 hidPs = zeros(sNum,hidNum);
wolffd@0 32 hidN = zeros(sNum,hidNum);
wolffd@0 33 hidNs = zeros(sNum,hidNum);
wolffd@0 34 %% Reconstruction error & evaluation error & early stopping
wolffd@0 35 mse = 0;
wolffd@0 36 omse = 0;
wolffd@0 37 inc_count = 0;
wolffd@0 38 MAX_INC = 3000; % If the error increase MAX_INC times continuously, then stop training
wolffd@0 39 %% Average best settings
wolffd@0 40 n_best = 1;
wolffd@0 41 aW = size(W);
wolffd@0 42 aVB = size(visB);
wolffd@0 43 aHB = size(hidB);
wolffd@0 44 %% Plotting
wolffd@0 45 h = plot(nan);
wolffd@0 46 %% ==================== Start training =========================== %%
wolffd@0 47 for i=1:conf.eNum
wolffd@0 48 if i== N+1
wolffd@0 49 lr = conf.params(2);
wolffd@0 50 end
wolffd@0 51 omse = mse;
wolffd@0 52 mse = 0;
wolffd@0 53 for j=1:conf.bNum
wolffd@0 54 visP = data((j-1)*conf.sNum+1:j*conf.sNum,:);
wolffd@0 55 %up
wolffd@0 56 hidP = logistic(visP*W + repmat(hidB,sNum,1));
wolffd@0 57 hidPs = 1*(hidP >rand(sNum,hidNum));
wolffd@0 58 hidNs = hidPs;
wolffd@0 59 for k=1:conf.gNum
wolffd@0 60 % down
wolffd@0 61 visN = logistic(hidNs*W' + repmat(visB,sNum,1));
wolffd@0 62 visNs = 1*(visN>rand(sNum,visNum));
wolffd@0 63 if j==5 && k==1, observe_reconstruction(visN,sNum,i,28,28); end
wolffd@0 64 % up
wolffd@0 65 hidN = logistic(visNs*W + repmat(hidB,sNum,1));
wolffd@0 66 hidNs = 1*(hidN>rand(sNum,hidNum));
wolffd@0 67 end
wolffd@0 68 % Compute MSE for reconstruction
wolffd@0 69 rdiff = (visP - visN);
wolffd@0 70 mse = mse + sum(sum(rdiff.*rdiff))/(sNum*visNum);
wolffd@0 71 % Update W,visB,hidB
wolffd@0 72 diff = (visP'*hidP - visNs'*hidN)/sNum;
wolffd@0 73 DW = lr*(diff - conf.params(4)*W) + conf.params(3)*DW;
wolffd@0 74 W = W + DW;
wolffd@0 75 % W = W.*mW;
wolffd@0 76 DVB = lr*sum(visP - visN,1)/sNum + conf.params(3)*DVB;
wolffd@0 77 visB = visB + DVB;
wolffd@0 78 DHB = lr*sum(hidP - hidN,1)/sNum + conf.params(3)*DHB;
wolffd@0 79 hidB = hidB + DHB;
wolffd@0 80 end
wolffd@0 81 %%
wolffd@0 82 mse_plot(i) = mse;
wolffd@0 83 axis([0 (conf.eNum+1) 0 2]);
wolffd@0 84 set(h,'YData',mse_plot);
wolffd@0 85 drawnow;
wolffd@0 86 % plot(mse_plot,'XDataSource','real(mse_plot)','YDataSource','imag(mse_plot)')
wolffd@0 87 % linkdata on;
wolffd@0 88
wolffd@0 89 if mse > omse
wolffd@0 90 inc_count = inc_count + 1
wolffd@0 91 else
wolffd@0 92 inc_count = 0;
wolffd@0 93 end
wolffd@0 94 if inc_count> MAX_INC, break; end;
wolffd@0 95 fprintf('Epoch %d : MSE = %f\n',i,mse);
wolffd@0 96 end
wolffd@0 97 end