Negative TrainedVariance error when running matlab example of image GAN(i change the code enlarge the image size from 64x64 to 128x128)

I follow this: https://www.mathworks.com/matlabcentral/answers/585146-how-to-change-gan-example-to-generate-images-with-a-larger-size to change the image size processed in GAN from 64x64 to 128x128. A guy said he can make it at the end of post but I got a negative TrainedVariance error when running it.

Here is the system error warning:

Error using nnet.internal.cnn.dlnetwork/set.State
Layer 'batchnorm_1': Invalid State. Expected TrainedVariance to be positive.

Error in dlnetwork/set.State (line 551)
net.PrivateNetworkStorage.State = values;

Error in trainGAN (line 70)
netG.State = stateG;

The error pumped out when it runs for 1 to 2 minutes.

I Googled it and tried to add those codes to replace the line netG.State = stateG;:

idx = netG.State.Parameter == "TrainedVariance";
boundAwayFromZero = @(X) max(X, eps('single'));
netG.State(idx,:) = dlupdate(boundAwayFromZero, netG.State(idx,:));

But cant work since netG.State isnt get value from stateG

[Ziqi Sun]

I add theses lines that bound TrianedVarience from zero:

        stateT=stateG;
        idx = stateT.Parameter == "TrainedVariance";
        boundAwayFromZero = @(X) max(X, eps('single'));
        stateT(idx,:) = dlupdate(boundAwayFromZero, stateT(idx,:));
        netG.State = stateT;

But then the generator can no longer generate anything out, probabaly because the TrianedVarience keeps drop to zero even if i prevent it drop to negative values...

By "no longer generate anything out", I mean after several epoches, the ouput of generator "I" when I image it, it is a blank...

[Suhail Mahmud]

Hi @Ziqi Sun, were you able to resolve the trainedVariance issue and generate images higher than 128 pixel?

[Ziqi Sun]

@Suhail Mahmud Unfortunately, no.

Answer 1

[Ayush Modi]

Hi Ziqi,

As per my understanding, you are trying to process 128x128 images in GAN but you are getting negative TrainedVariance error while running it.

I was able to achieve this by following the answer provided by @Fred Liu in the Matlab community question you are following (https://www.mathworks.com/matlabcentral/answers/585146-how-to-change-gan-example-to-generate-images-with-a-larger-size).

Below is the code for your reference:

filterSize = 5;
numFilters = 128;
numLatentInputs = 100;
projectionSize = [4 4 512];

layersGenerator = [
    featureInputLayer(numLatentInputs)
    projectAndReshapeLayer(projectionSize)
    transposedConv2dLayer(filterSize,8*numFilters)
    batchNormalizationLayer
    reluLayer
    transposedConv2dLayer(filterSize,4*numFilters,Stride=2,Cropping="same")
    batchNormalizationLayer
    reluLayer
    transposedConv2dLayer(filterSize,2*numFilters,Stride=2,Cropping="same")
    batchNormalizationLayer
    reluLayer
    transposedConv2dLayer(filterSize,numFilters,Stride=2,Cropping="same")
    batchNormalizationLayer
    reluLayer
    transposedConv2dLayer(filterSize,3,Stride=2,Cropping="same")
    tanhLayer];
netG = dlnetwork(layersGenerator);


dropoutProb = 0.5;
numFilters = 128;
scale = 0.2;
inputSize = [128 128 3];
filterSize = 5;

layersDiscriminator = [
    imageInputLayer(inputSize,Normalization="none")
    dropoutLayer(dropoutProb)
    convolution2dLayer(filterSize,numFilters,Stride=2,Padding="same")
    leakyReluLayer(scale)
    dropoutLayer(dropoutProb)
    convolution2dLayer(filterSize,2*numFilters,Stride=2,Padding="same")
    batchNormalizationLayer
    leakyReluLayer(scale)
    dropoutLayer(dropoutProb)
    convolution2dLayer(filterSize,4*numFilters,Stride=2,Padding="same")
    batchNormalizationLayer
    leakyReluLayer(scale)
    dropoutLayer(dropoutProb)
    convolution2dLayer(filterSize,8*numFilters,Stride=2,Padding="same")
    batchNormalizationLayer
    leakyReluLayer(scale)
    convolution2dLayer(8,1)
    sigmoidLayer];
netD = dlnetwork(layersDiscriminator);

I hope this helps!