Forward outputs from discriminator of GAN vary with the number of inputs.

I'm using MATLAB R2019b, and deep learning toolbox.

I wanna see sigmoid value of outputs from discriminator, but it vary with the number of inputs in dlarray.

For example,

I assigned input data X (size 1024,1,1,3) to dlX which is dlarray(X, 'SSCB')

--- dlX=dlarray(X,'SSCB');

so prediction can be extracted by calling function dlYPred=forward(dlnetDiscriminator, dlX), and out=sigmoid(dlYPred)

--- dlYPred=forward(dlnetDiscriminator, dlX);

--- out=sigmoid(dlYPred);

I got output from this process. -> (0.1, 0.3, 0.5)

But if i assign input data X (size 1024,1,1,5) that includes previous three data, results is changed.

like (0.2 0.4 0.7 0.3 0.1) although first three output should not be changed.

I wanna get data like (0.1 0.3 0.5 0.2 0.02)

How can i solve it?

Answer 1

[Sourav Bairagya]

It seems that the discriminator layer weights get chnaged during second time calling of "forward" function. Make sure that the discriminator layer weights are kept fixed while running "foward" function for second time.

You can also use "predict" function to compute the responses out of the trained network.

https://www.mathworks.com/help/deeplearning/ref/predict.html?s_tid=doc_ta