Combine Bayesian regularization (trainbr) training algorithm with weight regularization.

Question

Arygianni Valentino 2018-3-30

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/391594-combine-bayesian-regularization-trainbr-training-algorithm-with-weight-regularization

评论： Arygianni Valentino 2018-3-30

Hi everyone.

I want to ask is it okay to train a neural network using trainbr algorithm combine with network regularization in matlab? I try to use this idea, and the result is my training always stop due to reach mu_max. If I used trainbr only the train will stop due to max epoch or max validation fail.

Here is the code that I used:

TOL=0;
net = network;
net.numInputs = 1;
net.numLayers = 6;
net.biasConnect = [1;1;1;1;1;1];
net.inputConnect = [1;1;0;0;0;0];
net.layerConnect = [0 0 0 0 0 0;
                    0 0 0 0 0 0;
                    1 1 0 0 0 0;
                    0 0 1 0 0 0;
                    0 0 1 0 0 0;
                    0 0 0 1 1 0];
net.outputConnect = [0 0 0 0 0 1];
net.targetConnect = [0 0 0 0 0 1];
net.inputs{1}.range = ones(NID,2);
net.inputs{1}.range(:,1) = -1;
net.layers{1}.size = NNIL;
net.layers{1}.transferFcn = TFIL1;
net.layers{1}.initFcn = 'initnw';
net.layers{2}.size = NNIL;
net.layers{2}.transferFcn = TFIL2;
net.layers{2}.initFcn = 'initnw';
net.layers{3}.size = NLPC;
net.layers{3}.transferFcn = TFIL3;
net.layers{3}.initFcn = 'initnw';
net.layers{4}.size = NNIL;
net.layers{4}.transferFcn = TFIL1;
net.layers{4}.initFcn = 'initnw';
net.layers{5}.size = NNIL;
net.layers{5}.transferFcn = TFIL2;
net.layers{5}.initFcn = 'initnw';
net.layers{6}.size = NID;
net.layers{6}.transferFcn = TFIL3;
net.layers{6}.initFcn = 'initnw';
net.initFcn = 'initlay';
net.performFcn = 'mse';
net.performParam.regularization = 0.000001;
net.trainFcn = 'trainbr';
net.trainParam.epochs = 2000;
net.trainParam.min_grad = 0;
net.trainParam.minstep = 0;
net.trainParam.max_fail = 200;
net=init(net);
net.trainParam.goal = TOL;
net.divideFcn = 'divideind';
net.divideParam.trainInd = TrainIdx;    
net.divideParam.valInd = ValIdx;
net.divideParam.testInd = TestIdx;
[net,tr,Y,E,Pf,Af] = train(net,Data_All,Data_All);