train_noise_snn.m

来自「神经网络的工具箱, 神经网络的工具箱,」· M 代码 · 共 74 行

function [noise_net, noise_record] = train_noise_snn(net, noise_data, blocksize)
%TRAIN_NOISE_SNN Train a network on noise prediction.
%
% Syntax
%
%  [noise_net, noise_tr_info] = train_noise_snn(net, noise_data)
%
% Description
%
%  TRAIN_NOISE_SNN takes the connection weights, biases and transfer
%  functions of 'net' and replaces the transfer function of the last layer 
%  with an exponential transfer function. Then in a training
%  procedure, only the connection weights and biases in this last
%  layer are adjusted to minimize the cost on 'noise_data'.
%
%  TRAIN_NOISE_SNN(net, noise_data)
%    net         - a net_struct
%    noise_data  - a wcfdata_struct with a training data set for the
%                  noise.
%  and returns 
%    noise_net     - a net_struct predicting noise.
%    noise_tr_info - a struct containing info on the training process.
%
% See also
%
%  C_PREDICTION_SNN
%

%#function lintf_snn exptf_snn logsigtf_snn radbastf_snn tansigtf_snn 

if (nargin < 3)
   blocksize = 1;
end

L = net.numLayers;
MU = size(noise_data.P,2);

yh = feval(net.transferFcn{1}, net.weights{1}*noise_data.P + ...
                               repmat(net.biases{1}, 1, MU));
if (L>1)
   for l = 2:(L-1)
       yh = feval(net.transferFcn{l}, net.weights{l}*yh + ...
                                      repmat(net.biases{l},1,MU));
   end
end
noise_data.P = yh;

N = size(net.biases{net.numLayers},1);
Nh = size(net.biases{net.numLayers-1},1);

train_net = net_struct_snn([Nh N], {'exptf_snn'}, 'traincgp_snn');
train_net.costFcn.name = 'wcf_snn';
train_net.costFcn.fn = 'loglikelihood_snn';
train_net.costFcn.a = [];
if isfield(noise_data, 'useT')
   for i = 1:N
       mu = find(~isnan(noise_data.T(i,:)));
       NR = size(mu, 2);
       train_net.biases{1}(i) = log(1/NR * sum(exp(noise_data.T(i,mu)), 2)); 
   end
else
   train_net.biases{1} = log(1/MU * sum(exp(noise_data.T), 2)); 
end
train_net.weights{1} = zeros(N, Nh); 
[train_net, noise_record] = train_bootstrap_snn(train_net, noise_data, blocksize);

noise_net = net;
noise_net.transferFcn{L} = 'exptf_snn';
%noise_net.transferFcn{L} = 'logsig';
noise_net.biases{L} = train_net.biases{1};
noise_net.weights{L} = train_net.weights{1};