update_everything.m

非线型因素分析matlab仿真程序包

function [sources, net, oldgrads, status] = update_everything(...
    sources, net, dcp_dsm, dcp_dsv, fs, params, dcp_dnetm, dcp_dnetv, ...
    data, oldc, status, oldgrads)
% UPDATE_EVERYTHING  Perform a line search to find optimal step length
%     and update all parameters

% Copyright (C) 1999-2004 Antti Honkela, Harri Valpola,
% and Xavier Giannakopoulos.
%
% This package comes with ABSOLUTELY NO WARRANTY; for details
% see License.txt in the program package.  This is free software,
% and you are welcome to redistribute it under certain conditions;
% see License.txt for details.

[sources, net, newc, restc, status] = update_variances(...
    sources, net, dcp_dsv, fs, params, dcp_dnetv, data, oldc, status);

if status.updatesrcs < 0,
  dc_dsm = zeros(size(dcp_dsm));
else
  dc_dsm = -dcp_dsm;
end

if status.updatenet < 0,
  dc_dnetm.w1 = zeros(size(dcp_dnetm.w1));
  dc_dnetm.b1 = zeros(size(dcp_dnetm.b1));
  dc_dnetm.w2 = zeros(size(dcp_dnetm.w2));
  dc_dnetm.b2 = zeros(size(dcp_dnetm.b2));
else
  dc_dnetm.w1 = -dcp_dnetm.w1;
  dc_dnetm.b1 = -dcp_dnetm.b1;
  dc_dnetm.w2 = -dcp_dnetm.w2;
  dc_dnetm.b2 = -dcp_dnetm.b2;
end

if strcmp(status.updatealg, 'conjgrad'),
  gn = sum(dc_dsm(:).^2) + ...
       sum(dcp_dnetm.w1(:).^2) + sum(dcp_dnetm.w2(:).^2) + ...
       sum(dcp_dnetm.b1(:).^2) + sum(dcp_dnetm.b2(:).^2);
  if (oldgrads.norm ~= 0) & all(size(oldgrads.s) == size(dc_dsm)) ...
	& all(size(oldgrads.net.w1) == size(dc_dnetm.w1)),
    beta = gn / oldgrads.norm;

    dc_dsm = dc_dsm + beta * oldgrads.s;
    dc_dnetm.w1 = dc_dnetm.w1 + beta * oldgrads.net.w1;
    dc_dnetm.w2 = dc_dnetm.w2 + beta * oldgrads.net.w2;
    dc_dnetm.b1 = dc_dnetm.b1 + beta * oldgrads.net.b1;
    dc_dnetm.b2 = dc_dnetm.b2 + beta * oldgrads.net.b2;
  end
  oldgrads.norm = gn;
  oldgrads.s = dc_dsm;
  oldgrads.net = dc_dnetm;
end

[sources, net, status.t0] = linesearch(...
    sources, net, newc, restc, dc_dsm, dc_dnetm, fs, data, params, ...
    status.t0, status);



function [sources, net, newc, restc, status] = update_variances(...
    sources0, net0, dcp_dsv, fs, params, dcp_dnetv, ...
    data, oldcost, status)
% UPDATE_VARIANCES  Update the variances of the network and the sources

epsilon=1e-6;
sources = sources0;
net = net0;

if ~isfield(status, 'varalpha'),
  status.varalpha = 1;
end

newc = inf;
alpha = min([1, 2*sqrt(status.varalpha)]);

if status.updatesrcvars >= 0,
  newsvar = .5 ./ max(dcp_dsv, .45 ./ sources.var);
end

if status.updatenet >= 0,
  neww1var = .5 ./ max(dcp_dnetv.w1, .45 ./ net.w1.var);
  neww2var = .5 ./ max(dcp_dnetv.w2, .45 ./ net.w2.var);
  newb1var = .5 ./ max(dcp_dnetv.b1, .45 ./ net.b1.var);
  newb2var = .5 ./ max(dcp_dnetv.b2, .45 ./ net.b2.var);
end

while newc > oldcost + epsilon,
  if status.updatesrcvars >= 0,
    sources.var = exp(alpha * log(newsvar) + (1-alpha) * log(sources0.var));
  end

  if status.updatenet >= 0,
    net.w1.var = exp(alpha * log(neww1var) + (1-alpha) * log(net0.w1.var));
    net.w2.var = exp(alpha * log(neww2var) + (1-alpha) * log(net0.w2.var));
    net.b1.var = exp(alpha * log(newb1var) + (1-alpha) * log(net0.b1.var));
    net.b2.var = exp(alpha * log(newb2var) + (1-alpha) * log(net0.b2.var));
  end
  
  fs_tmp = feedfw(sources, net, status.approximation);
  [netc, restc] = kl_static_split(net, params);
  newc = kl_batch(fs_tmp{4}, sources, data, params) + netc + restc;
  if status.debug,
    fprintf('Variance update: alpha=%f, c=%f\n', alpha, newc);
  end
  alpha = .5 * alpha;
end

if (2*alpha < min([1, 2*status.varalpha])) && status.debug,
  fprintf('Decreased variance step...\n');
end

status.varalpha = 2*alpha;