Want to use something like dividerand and then apply the indices over and over with various fitnet & train combinations?

Question

Brent 2018-12-22

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/436884-want-to-use-something-like-dividerand-and-then-apply-the-indices-over-and-over-with-various-fitnet

编辑： Brent 2018-12-23

For example in the following code after running dividerand I can see the indices for the train, validation and test values.

However as soon as it sees trainbr, it eliminates the validation indices, which I don't want it to do so. When I create a new fitnet using trainlm it is no longer using the same train, validation and test values.

Any ideas?

x = -1:0.05:1;
t = sin(2*pi*x) + 0.1*randn(size(x));
[trainT,valT,testT] = dividerand(length(t),80/100,10/100,10/100);  %  maybe make eval percentage 0 if trainbr, but not necessary since: 
                                                                   %  "Validation stops are disabled by default (max_fail = 0) so that
                                                                   %  training can continue until an optimial combination of errors and
                                                                   %  weights are found"?
net = fitnet([20,20],'trainbr');    % [20,10] means 2 hidden layers 1st w size 20 & 2nd w size 10
% help(net.trainFcn) to see the default settings
% disable fitnet from using mapminmax if using my scaling algo
[net,tr] = train(net,x,t);  % trains the NN and tr contains all kinds of metadata
view(net)    % sb after train
y = net(x);
perf = perform(net,y,t)
%doing fitnet then train again to see if still using same indices for
%trainT,valT, testT in the train function
net = fitnet([20,10],'trainlm');    % [20,10] means 2 hidden layers 1st w size 20 & 2nd w size 10
% help(net.trainFcn) to see the default settings
% disable fitnet from using mapminmax if using my scaling algo
[net,tr] = train(net,x,t);  % trains the NN and tr contains all kinds of metadata
view(net)    % sb after train
y = net(x);
perf = perform(net,y,t)