MATLAB: How to train an autoencoder without trainAutoencoder function

Question

I what train an autoencoder without trainAutoencoder function.

Then I got some problems.

The learning of the network stopped at a certain value, and the output image was different from the input.

Despite the nodes of the hidden layer being larger than the input size.

I must have overlooked, but I can not consider it.

The code and output.

rng('default')
% Load the training data into memory
[xTrainImages,tTrain] = digitTrainCellArrayData;
N = length(xTrainImages);
hiddenSize = 1000;
%%cell to array(alter cell into vector)
T = zeros(numel(xTrainImages),28*28);
for i = 1:numel(xTrainImages)
   T(i,:) = xTrainImages{i}(:)';
end
X = T';
T = T';
%%neural networks setting and training
net = fitnet(hiddenSize);
net.trainFcn = 'trainscg';
net = configure(net,X,T);
net = train(net,X,T);
%%denoising images
h1 = figure;
h2 = figure;
figure(h1)
for i = 1:20
    subplot(4,5,i);
    imshow(xTrainImages{i})
end
figure(h2);
for i = 1:20
    subplot(4,5,i);
    imshow(reshape(net(xTrainImages{i}(:)),28,28));
end

Answer 1

I solved it after sparse learning.