view toolboxes/RBM/gen_training_rbm.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
line wrap: on
line source
function [W visB hidB] = gen_training_krbm(conf,W,mW,train_file,train_label)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Training Knowledge Based RBM for generative classification         %  
% conf: training setting                                             %
% W: weights of connections                                          %
% mW: mask of connections                                            %
% -*-sontran2012-*-                                                  %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% load data
vars = whos('-file', train_file);
A = load(train_file,vars(1).name);
data = A.(vars(1).name);
vars = whos('-file', train_label);
A = load(train_label,vars(1).name);
label = A.(vars(1).name);
assert(~isempty(data),'[KRBM-GEN] Data is empty'); 
assert(size(data,1) == size(label,1),'[KRBM-GEN] Number of data and label mismatch'); 
Classes = unique(label)';
lNum = size(Classes,2);
data = [data discrete2softmax(label,Classes)]                              
%% initialization
visNum  = size(data,2);
hidNum  = conf.hidNum;
sNum  = conf.sNum;
lr    = conf.params(1);
N     = 10;                                                                     % Number of epoch training with lr_1                     
W     = [W;0.1*randn(visNum - size(W,1),size(W,2))];
W     = [W 0.1*randn(size(W,1),hidNum-size(W,2))];

DW    = zeros(size(W));
visB  = zeros(1,visNum);
DVB   = zeros(1,visNum);
hidB  = zeros(1,hidNum);
DHB   = zeros(1,hidNum);
visP  = zeros(sNum,visNum);
visN  = zeros(sNum,visNum);
visNs = zeros(sNum,visNum);
hidP  = zeros(sNum,hidNum);
hidPs = zeros(sNum,hidNum);
hidN  = zeros(sNum,hidNum);
hidNs = zeros(sNum,hidNum);
%% Reconstruction error & evaluation error & early stopping
mse    = 0;
omse   = 0;
inc_count = 0;
MAX_INC = 3;                                                                % If the error increase MAX_INC times continuously, then stop training
%% Average best settings
n_best  = 1;
aW  = size(W);
aVB = size(visB);
aHB = size(hidB);
%% ==================== Start training =========================== %%
for i=1:conf.eNum
    if i== N+1
        lr = conf.params(2);
    end
    omse = mse;
    mse = 0;
    for j=1:conf.bNum
       visP = data((j-1)*conf.sNum+1:j*conf.sNum,:);
       %up
       hidP = logistic(visP*W + repmat(hidB,sNum,1));
       hidPs =  1*(hidP >rand(sNum,hidNum));
       hidNs = hidPs;
       for k=1:conf.gNum
           % down
           visN  = hidNs*W' + repmat(visB,sNum,1);
           visN(:,1:visNum-lNum) = logistic(visN(:,1:visNum-lNum));
           visN(:,visNum-lNum+1:visNum) = softmax_activation(visN(:,visNum-lNum+1:visNum));
           visNs = [1*(visN(:,1:visNum-lNum)>rand(sNum,visNum-lNum)) visN(:,visNum-lNum+1:visNum)];
           if j==5 && k==1, observe_reconstruction(visN(:,1:visNum-lNum),sNum,i,28,28); end
           % up
           hidN  = logistic(visNs*W + repmat(hidB,sNum,1));
           hidNs = 1*(hidN>rand(sNum,hidNum));
       end
       % Compute MSE for reconstruction
       rdiff = (visP - visN);
       mse = mse + sum(sum(rdiff.*rdiff))/(sNum*visNum);
       % Update W,visB,hidB
       diff = (visP'*hidP - visNs'*hidN)/sNum;
       DW  = lr*(diff - conf.params(4)*W) +  conf.params(3)*DW;
       W   = W + DW;
%        W = W.*mW;
       DVB  = lr*sum(visP - visN,1)/sNum + conf.params(3)*DVB;
       visB = visB + DVB;
       DHB  = lr*sum(hidP - hidN,1)/sNum + conf.params(3)*DHB;
       hidB = hidB + DHB;
    end
    if mse > omse
        inc_count = inc_count + 1
    else
        inc_count = 0;
    end
    if inc_count> MAX_INC, break; end;
    fprintf('Epoch %d  : MSE = %f\n',i,mse);
end
end