Error using * in dotprod.apply when training custom network with train

Hi everyone,

I have a neural network that is initiated with the following code:

close all

clear all %#ok<CLALL>

clc

load iris_dataset; % Load iris dataset

% Modify irisTargets to include a fourth output that is always zero. This

% makes it compatible with a network that has two output 2x2 layers.

irisTargets=[irisTargets; zeros(1,length(irisTargets))];

net=network; % Generate network

net.numInputs=4; % 4 inputs

net.inputs{:}.size=1; % Input size of 1

net.numlayers=6; % 6 total layers

net.layers{1:4}.size=1; % First 4 layers are hidden, and

% have one neuron each.

net.layers{5:6}.size=2; % Last 2 layers are output layers,

% and have two neurons each.

net.inputconnect(1:2,1:2)=1; % First 4 layers are fully

% connected with 2 of the 4

% inputs. Simulates 2x2

% network, and can be 'split'

% later on.

net.inputconnect(3:4,3:4)=1; % Last 2 layers are fully connected

% with 2 of the 4 hidden

% layers.

net.inputweights{1:2,1:2}.initfcn='initzero'; % Initialize input weights to 0

net.inputweights{3:4,3:4}.initfcn='initzero';

net.inputweights{1:2,1:2}.learnfcn='learnp'; % Learn with input weights > 0 ?

net.inputweights{3:4,3:4}.learnfcn='learnp';

net.layerconnect(5,1:2:3)=1; % Connect layers 1 and 3 to

% layer 5

net.layerconnect(6,2:2:4)=1; % Connect layers 2 and 4 to

% layer 6

net.layers{1:4}.transferfcn='poslin'; % ReLU activation function

net.layers{5:6}.transferfcn='hardlim'; % Step activation function

net.layerweights{5,1:2:3}.initfcn='initzero'; % Initialize layer weights to 0

net.layerweights{6,2:2:4}.initfcn='initzero';

net.layerweights{5,1:2:3}.learnfcn='learnp'; % Learn withh layer weights > 0 ?

net.layerweights{6,2:2:4}.learnfcn='learnp';

net.outputconnect(5:6)=1; % Connect layers 5 and 6 to outputs

net.initfcn='initlay';

net.trainfcn='trainc';

net.performfcn='mae';

view(net)

net=train(net, irisInputs, irisTargets); % Train network with iris

% dataset inputs and

% modified iris dataset

% outputs. This is where

% the error occurs.

It generates the following output from view(net):

However, when the code reaches the last line, it generates the following error:

Error using *

Inner matrix dimensions must agree.

Error in dotprod.apply (line 6)

z = w*p;

Error in nn7.y_all

Error in nn7.perf_all (line 5)

data = nn7.y_all(net,data,fcns);

Error in nn7.perf_sig_grad (line 7)

[perf,data] = nn7.perf_all(net,data,fcns);

Error in trainc>train_network (line 198)

[dperf,divData,gB,gIW,gLW,gA] = nn7.perf_sig_grad(net,divData,needGradient,fcns);

Error in trainc (line 55)

[out1,out2] = train_network(varargin{2:end});

Error in network/train (line 353)

[net,tr] = feval(net.trainFcn,'apply',net,tr,data,calcMode.options,hints,net.trainParam);

Error in mynn (line 58)

net=train(net, irisInputs, irisTargets); % Train network with iris

I have done as much digging as I can, considering the limited access I have to the files, and I found that, at some point, the "p" input to "dotprod.apply" became empty. That said, I cannot figure out why, since the file that calls "dotprod.apply" has restricted access. I did make sure to double check the matrix calculations that should be going on before actually running the code, so I am really confused. Any help with this would be greatly appreciated!

Cheers!