CNN inputs and outputs are incorrect and I don't know why

Hello, I've been trying to set up a CNN to predict the previous state in a Conway's Game of Life. Basically I've set up random boards with associated next step to train my neural network. The batchcreattion function simply creates board and run the game for one iteration. In that case I get a NxNxK array.

N = 20;
K = 10000;
[X,Y] = batchCreation(K,N);
layers = [
    % Setup the inputs
    imageInputLayer([N N 1])
    
    % Five 5x5 filters
    convolution2dLayer(5,5,'Padding','same')
    batchNormalizationLayer 
    reluLayer
    
    % Connect the output to the network
    fullyConnectedLayer(N*N)
    softmaxLayer % The results from the fully connected layer are converted to explicit probabilities by the soft max - it normalizes them to between 0 and 1
    classificationLayer];

Xvalid = X(:,:,1:K/2);
Yvalid = Y(:,:,1:K/2);
options = trainingOptions('sgdm', ...
    'InitialLearnRate',0.01, ...
    'MaxEpochs',500, ...
    'Shuffle','every-epoch', ...
    'ValidationData',{Xvalid,Yvalid}, ...
    'Plots','training-progress');
Xtrain = X(:,:,K/2:K); 
Ytrain = Y(:,:,K/2:K);
net = trainNetwork(Xtrain,Ytrain,layers,options);

But for some reason I get the following message: `Error using trainNetwork Invalid training data. For classification tasks, responses must be a vector of categorical responses. For regression tasks, responses must be a vector, a matrix, or a 4-D array of numeric responses which must not contain NaNs.`

I'm really confused by how to set it up properly. Any help would be appreciated!

Answer 1

[Walter Roberson]

K = 10000;
[X,Y] = batchCreation(K,N);

We do not have that function, batchCreation, so we do not know at the moment what size(X) and size(Y) are.

Xtrain = X(:,:,K/2:K); 
Ytrain = Y(:,:,K/2:K);

but we can see there that slightly over half of the third dimension is being selected (half would be K/2+1:K) . The result cannot possibly be a vector, because "vector" for the purposes of trainNetwork must be either a row vector or column vector.

The structure of the code hints that Y probably is not scalar in one of the first two dimensions, so Y(:,:,K/2:K) is probably either a full 3D or else has probably as two non-scalar dimensions. If it happens to be 1 x 1 x K then you should squeeze(Y(:,:,K/2:K))

[Chris]

Hello Walter, I precised the size of my arrays in my post,. [X,Y] are two arrays of size NxNxK. I basically have a set of K boards of size NxN. But for more details here are my two other functions:

Conway's game of life:

function [outputBoard] = Conway(Board)
%CONWAY Calculate next step in Conway's Game of Life
%   Input: Board, array
%   Output: Board after one iteration of the game, array
t = conv2(Board(:,:),[2,2,2;2,1,2;2,2,2],'same');    % simple convolution over the matrix
outputBoard = (t > 4) & (t < 8);                % the new value is a boolean with value True or False, or, 1 or 0. We apply conway's rules to set the new valu
end

And the batch creation:

function [X,Y] = batchCreation(K,N)
%batchCreation creates batch of inputs and associated outputs to be used in
%the training of the neural network
%   K: number of boards, int
%   N: size of board, int
p=0.5; % Bernouilli distribution

Y = rand(N,N,K);
Y =(Y<p);

X = zeros(N,N,K);
for j=1:K
    X(:,:,j) = Conway(Y(:,:,j));
end
end

[Walter Roberson]

If you look at the possibilities in https://www.mathworks.com/help/deeplearning/ref/trainnetwork.html#mw_d0b3a2e4-09a0-42f9-a273-2bb25956fe66 then you will see that the only time you could have a 3D array of responses is if the task is 2D image regression and the response variable is height by width by channels by number_of_images -- so for 3D response variable it would have to be the case that you had K/2+1 channels and only one image. If you wanted height by width by 1 channel x N images then you would need to permute(Ytrain, [1 2 4 3])

[Chris]

What I want to do is train my network with K images of size NxN. So what would you suggest because I'm really lost in the documentation at the moment.

[Walter Roberson]

permute(Ytrain, [1 2 4 3])

[Chris]

Well I am getting the same error `Invalid training data. For classification tasks, responses must be a vector of categorical responses. For regression tasks, responses must be a vector, a matrix, or a 4-D array of numeric responses which must not contain NaNs.`