Testing Neural Network on new data

I am quite new to MATLAB and extrmely new to neural networks. I have created a feed forward neural network using newff(). I input a matrix of size 486*1200 and output matrix 6*1200. When I simulate the network using sim() function after training, I get correct output. But when I try to simulate on any other input of shorter size I get error which says input matrix should have 486 rows (as that of training matrix).

Can anyone resolve this issue for me to generalize the network for any size of input? And if I make that matrix of 486 rows by padding extra zeros, it gives incorrect output even though that test matrix is a part of trainig matrix.

Here is my code:

[Pr,Pc] = size(PP);% Pr=486, Pc=1200

TP=eye(6,1200);
[Tr,Tc] = size(TP);
L1 = 0.5*Pc;

net0to9 = newff(minmax(PP),[L1 Tr],{'logsig' 'logsig'},'traingdx');

net0to9.performFcn = 'mse';        
net0to9.trainParam.goal = 0;     
net0to9.trainParam.show = 20;      
net0to9.trainParam.epochs =1200;
[net0to9,Tr] = train(net0to9,PP,TP);

x = sim(net0to9,PP);

I'd be really thankful if anyone could help.

[Greg Heath]

Once a net with node topology I-H-O is trained,

It should only be used with input matrices with

I = size(input,1)

Inputs with size

[I Ntst] = size(input)

will yield output matrices of size

[O Ntst] = size(output)

Hope this helps.

Greg

P.S. It is assumed that nontraining data has

the same statistical characteristics as the training data.

Answer 1

[Walter Roberson]

http://www.mathworks.com/help/releases/R2010a/toolbox/nnet/newff.html

For newff() the samples run down the columns!

Note the documentation says:

P  R x Q1 matrix of Q1 sample R-element input vectors

So the P(:,1) is the first sample, P(:,2) is the second sample, and so on.

Therefore, any data you test against must have Pr rows or else the samples will be incomplete.

If you intend that your samples are incomplete, then in order for a meaningful computation to take place, you would somehow have to indicate the mapping between the inputs you have available and the original input positions. I do not know if there is any mechanism for that at all, but if I were designing such a mechanism, the way I would probably expect the user to indicate missing inputs would be by putting NaN in that input location.

I do not know much about NN, but I have seen several people run in to this same problem of not noticing that the samples run down the columns.

[Bushra]

Thanx Walter

I am trying to reshape my input and output accordingly. It will work i guess. Thanx again for help.

Answer 2

[Greg Heath]

Here is my code:

>[Pr,Pc] = size(PP);% Pr=486, Pc=1200

Pr is probably much larger than necessary. Try input variable subset selection

using PCA (regression) or PLS (classification).

Also standardize the inputs to have zero-mean/unit-variance

>TP=eye(6,1200);

This makes no sense because

eye(6,1200) = [ eye(6), zeros(6, 1194) ]

Please explain your output. Is the problem regression or classification?

>[Tr,Tc] = size(TP);

>L1 = 0.5*Pc;

L1 = 600 is a ridiculous number of hidden nodes.

For an I-H-O net topology with [I N] = size(PP) and [O N] = size(TP), there are Neq = N*O = 1200*6 = 7200 training equations to estimate Nw = (I+1)*H+(H+1)*O = O+(I+O+1)*H weights. For accurate estimates, it is desired that Neq >> Nw or

H << (Neq-O)/(I+O+1) = 7194/493 = 14.6

Search the newsgroup using

heath Neq Nw

for details.

>net0to9 = newff(minmax(PP),[L1 Tr],{'logsig' 'logsig'},'traingdx'); >net0to9.performFcn = 'mse'; >net0to9.trainParam.goal = 0; >net0to9.trainParam.epochs =1200; >[net0to9,Tr] = train(net0to9,PP,TP); >x = sim(net0to9,PP);

1. Why not use as many defaults as is practical? See help newff and doc newff.

2. Why not use a more practical goal? Search the newsgroup using

heath net.trainParam.goal

3. Why not use additional outputs in train to obtain outputs and errors?

Consequently

net0to9 = newff(minmax(PP),[ H O],{'tansig' 'logsig'}); % H <= 14, O =6

net0to9.trainParam.goal = mean(var(TP))/100; % R^2 ~ 0.99

net0to9.trainParam.show = 20;

[net0to9 tr Y E ] = train(net0to9,PP,TP);

>I'd be really thankful if anyone could help.

Hope this helps.

Greg

[Bushra]

Thanx Greg,

My concept is much cleared now. I am trying to work out and will be back if faced any problem. Thanx alot for help.