How to add 2 or more hidden layer to the neural network?

I have this code i need to add 2 hidden layer, can anyone please help me with that please. Here is the code

[J,grad] = netcost(nn_params,input_layer_size,hidden_layer_size,num_labels,X, y, lambda)

Theta1 = reshape(nn_params(1:hidden_layer_size * (input_layer_size + 1)), ... hidden_layer_size, (input_layer_size + 1));

Theta2 = reshape(nn_params((1 + (hidden_layer_size * (input_layer_size + 1))):end), ... num_labels, (hidden_layer_size + 1));

num_labels=2; 1 hidden layer; input neurons=2160,hidden layer neuron=20; I want to add totally 2 hidden layers, please can anyone help me with the code please

[Greg Heath]

What are the dimensions of your input and target matrices?

A MLP with 1 hidden layer is a universal approximator.

What is your justification for wanting 2 hidden layers?

Greg

[Tousif Ahmed]

input and target images containing faces, having the size of 27x18 for training and for test having the size of 150x65; I need to add the hidden layer so that i can tabulate the variation in the result when 1 hidden layer and when more than 1 is used

[Greg Heath]

You will have to explain your problem in more detail including copious comments in the code. In particular, please explain

1. the difference in sizes for training and test subsets

2. Why no validation subset?

[Greg Heath]

In the MATLAB NN Toolbox, N I-dimensional "I"nput vectors are paired with N O-dimensional "O"utput target vectors. The resulting input and target matrices have the sizes

 [ I N ] = size(input)

 [O N ]  = size(target)

What are I, O and N in your problem ???

Greg

[Tousif Ahmed]

from the point what i understood from your question, training(input) image of size 27x18, having 100 images and target image of "any size". Is this what you are asking? Can you please tel about vector in detail? Please dont mind as i am new to matlab.

[Greg Heath]

I am asking you to use the size command to make sure your input and target matrices have the correct size. The actual size of the image is, typically, irrelevant because input vectors usually consist of extracted image features placed into column vectors.

Hope this helps.

Greg

[Tousif Ahmed]

yes, thank You. how can i select the number of neurons required for the operation? hiddenlayer size i have given as 100, inputs are 20, totally 2 hidden layer i am using. feedforwardnet([50 50]) i am using this, but when i am increasing above 100 its showing error

[Philip G]

Just executing the command

 net=feedforwardnet([200 200 200]);

Does not give any errors and will give you a network with 3 hidden layers of size 200 (see attached image). I am not sure what your error is. Also in the code above "netcost" is not a MATLAB function - so it is hard to say where your error comes from (especially this strange behavior that you can not create a network with a layer size bigger than 100).

[Tousif Ahmed]

i will post the code here. please have a look.

load latest.mat; trainFcn = 'trainscg'; % Scaled conjugate gradient backpropagation.

net.performFcn = 'mse'; %newly added

hiddenLayerSize = 10; %net = patternnet(hiddenLayerSize); net=feedforwardnet([50 10]);

view (net); net.divideParam.trainRatio = 70/100; net.divideParam.valRatio = 15/100; net.divideParam.testRatio = 15/100;

net.trainParam.epochs = 100;

[net,tr] = train(net,training,test); y = net(training); e = gsubtract(test,y); performance = perform(net,test,y) tind = vec2ind(test); yind = vec2ind(y); percentErrors = sum(tind ~= yind)/numel(tind);

net.IW{1,1}; net.b{1}; %net.LW {1};

% View the Network view(net)

 I have attached the file named 'latest.mat'.It contains image file of size 27x18, 11 images for training and 2 images for test. I dont know whether the procedure i have followed to add the images for training and test are correct. Please correct me if i am wrong and please have a look