trainlm_snn.m

来自「神经网络的工具箱, 神经网络的工具箱,」· M 代码 · 共 217 行

function [net, result] = trainlm_snn(net, dataLV, dataVV, dataTV)
%TRAINLM_SNN Levenberg-Marquardt backpropagation.
%
%  Syntax
%
%   [net, tr_info] = trainlm_snn(net, dataLV)
%   [net, tr_info] = trainlm_snn(net, dataLV, dataVV)
%   [net, tr_info] = trainlm_snn(net, dataLV, dataVV, dataTV)
%
%   trainfcn_struct = trainlm_snn('pdefaults')
%
%  Description
%
%   TRAINLM_SNN is a network training function that updates weight and
%   bias values according to Levenberg-Marquardt optimization.
%   
%   TRAINLM_SNN takes
%     net    - a net_struct containing the initial network.
%     dataLV - training data set.
%     dataVV - validation data set (optional).
%     dataTV - test data set (optional).
%   and returns 
%     net     - the trained network.
%     tr_info - a structure containing information about the
%               training process.
%
%   The field 'trainFcn' of 'net' must contain a structure with the
%   training parameters. The default parameters can be set with
%
%     net.trainFcn = trainlm_snn('pdefaults')
%
%   or can be set manually. The parameters are:
%
%     net.trainFcn.name     - 'trainlm_snn'
%     net.trainFcn.epochs   - maximum number of training epochs 
%     net.trainFcn.goal     - training goal for cost function 
%     net.trainFcn.max_fail - maximum number of epochs in which cost 
%                             on validation set increases.
%     net.trainFcn.max_suc_fail - maximum number of succesive epochs
%                                 in which cost on validation set increases.
%     net.trainFcn.mu       - initial mu
%     net.trainFcn.mu_dec   - factor for decreasing mu
%     net.trainFcn.mu_inc   - factor for increasing mu
%     net.trainFcn.mu_max   - maximum for mu
%     net.trainFcn.progressFcn - progressfcn_struct containing parameters
%                                for showing progress of training
%     net.trainFcn.time     - maximum time for training.
%
%  See also
%
%   NET_STRUCT_SNN, PROGRESSFCN_STRUCT_SNN 
%
%

% FUNCTION INFO
% =============

if isstr(net)
  switch (net)
    case 'pnames',
          net = fieldnames(trainlm_snn('pdefaults'));
    case 'pdefaults',
	  trainFcn.name = 'trainlm_snn';
          trainFcn.epochs = 1000;
          trainFcn.goal = 0;
          trainFcn.max_fail = 20;
          trainFcn.max_suc_fail = 10;
	  trainFcn.mem_reduc = 1;
	  trainFcn.mu = 0.001;
	  trainFcn.mu_dec = 0.1;
	  trainFcn.mu_inc = 10;
	  trainFcn.mu_max = 1e10;
          trainFcn.progressFcn = progress_snn('pdefaults');
          trainFcn.time = inf;
          net = trainFcn;
    otherwise,
          error('Unrecognized code.')
  end   
  return
end     

        
% CALCULATION
% ===========

data = dataLV;

if (nargin <3)
   doValidation = 0;
else
   doValidation = 1;
   data = [data dataVV];
end

if (nargin <4)
   doTest = 0;
else
   doTest = 1;
   data = [data dataTV];
end

% Constants
epochs = net.trainFcn.epochs;
goal = net.trainFcn.goal;
max_fail = net.trainFcn.max_fail;
max_suc_fail = net.trainFcn.max_suc_fail;
%mem_reduc = net.trainFcn.mem_reduc;
mu = net.trainFcn.mu;
mu_inc = net.trainFcn.mu_inc;
mu_dec = net.trainFcn.mu_dec;
mu_max = net.trainFcn.mu_max;
time = net.trainFcn.time;

% Initialize
result.stop = '';
startTime = clock;
X = getx_snn(net);
numParameters = length(X);
ii = sparse(1:numParameters, 1:numParameters, ones(1, numParameters));
result.tr = tr_struct_snn(epochs, 'perf', 'vperf', 'tperf', 'mu');
%#function wcf_snn
result.tr.perf(1) = feval(net.costFcn.name, net, dataLV);
result.tr.mu(1) = mu;
if (doValidation)
   result.best_X = X;
   result.tr.vperf(1) = feval(net.costFcn.name, net, dataVV);
   result.vperf_min = result.tr.vperf(1);
   result.num_fail = 0;
   result.num_suc_fail = 0;
end
if (doTest)
   result.tr.tperf(1) = feval(net.costFcn.name, net, dataTV);
end

% Train
for epoch = 1:epochs
 
    % Train with Levenberg Marquardt
    [jj, je] = fisherff_snn(X, net, dataLV);
    while (mu <= mu_max)
        dX = -(jj+ii*mu) \ je;
        X2 = X + dX;
        net2 = setx_snn(net,X2);
        perf2 = feval(net.costFcn.name, net2, dataLV);
        if (perf2 < result.tr.perf(epoch))
          X = X2;
	  perf = perf2;
          mu = mu * mu_dec;
          break
        end
        mu = mu * mu_inc;
    end
    if (mu> mu_max)
       result.stop = 'Maximum MU reached, performance goal was not met.';
       perf = result.tr.perf(epoch);
    end
    net = setx_snn(net, X); 

    % Save performance
    epochPlus1 = epoch + 1;
    result.tr.perf(epochPlus1) = perf;
    result.tr.mu(epochPlus1) = mu;
        
    % Validation
    if (doValidation)
       result.tr.vperf(epochPlus1) = feval(net.costFcn.name, net, dataVV);
       if (result.tr.vperf(epochPlus1) < result.vperf_min)
          result.vperf_min = result.tr.vperf(epochPlus1);
          result.best_X = X;
          result.num_fail = 0;
       elseif (result.tr.vperf(epochPlus1) > result.vperf_min)
          result.num_fail = result.num_fail + 1;
       end
       if (result.tr.vperf(epochPlus1) > result.tr.vperf(epoch))
          result.num_suc_fail = result.num_suc_fail + 1;
       else
          result.num_suc_fail = 0;
       end
    end
    if (doTest)
       result.tr.tperf(epochPlus1) = feval(net.costFcn.name, net, dataTV);
    end
    
    % Stopping criteria
    result.runtime = etime(clock, startTime);
    if (result.tr.perf(epochPlus1) <= goal)
       result.stop = 'Performance goal met.';
    elseif (epoch == epochs)
       result.stop = 'Maximum epoch reached, performance goal was not met.';
    elseif (result.runtime > time)
       result.stop = 'Maximum time elapsed, performance goal was not met.';
    elseif (doValidation) & (result.num_fail > max_fail)
       result.stop = 'Validation stop; maximum total failures';
       net = setx_snn(net, result.best_X);
    elseif (doValidation) & (result.num_suc_fail > max_suc_fail)
       result.stop = 'Validation stop; maximum successive failures';
       net = setx_snn(net, result.best_X);
    elseif (doValidation) & (mu> mu_max)
       net = setx_snn(net, result.best_X);
    end
    result.epoch = epoch;
    if length(result.stop) 
       %#function progress_snn
       feval(net.trainFcn.progressFcn.name, net, data, result);
       stdout_snn('%s, %s\n', upper(net.trainFcn.name), result.stop); 
       break; 
    end

    % Show progress
    %#function progress_snn
    feval(net.trainFcn.progressFcn.name, net, data, result);
end

% Finish
result.tr = cliptr_snn(result.tr, epoch);