How to change architecture of conditional GAN to generate 224x224x3 images?

2 次查看（过去 30 天）
Alok
Alok 2022-8-4
编辑： Alok 2022-8-6
I am following matlab example on conditional GAN at https://www.mathworks.com/help/deeplearning/ug/train-conditional-generative-adversarial-network.html
This example is for image size 64x64x3. I am wondering what changes should be done in layersGenerator and layersDiscriminator to generate 224x224x3 images.
This is my code:
inputSize = [224 224 3] or [256 256 3];
Note if Factor=2 (below) then I get image size 128x128x3. If Factor=4 then generated image size if 256x256x3. However, during the loop, it gives an error that trainedVariance is negative.
inputSize = [64 64 3];
Factor = 4; %if Factor =2 then 128x128x3 image size is generated;
inputSize = Factor*inputSize(1:2);
numClasses = 2;
augimds = augmentedImageDatastore(inputSize(1:2),XTrain,YTrain);
augimdsValidation = augmentedImageDatastore(inputSize(1:2),XValidation,YValidation);
numLatentInputs = 100;%100
embeddingDimension = 50;
numFilters = Factor*64;%224;
filterSize = 5;
projectionSize = Factor*[4 4 1024];
layersGenerator = [
featureInputLayer(numLatentInputs)
fullyConnectedLayer(prod(projectionSize))
functionLayer(@(X) feature2image(X,projectionSize),Formattable=true)
concatenationLayer(3,2,Name="cat");
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];
lgraphGenerator = layerGraph(layersGenerator);
layers = [
featureInputLayer(1)
embeddingLayer(embeddingDimension,numClasses)
fullyConnectedLayer(prod(projectionSize(1:2)))
functionLayer(@(X) feature2image(X,[projectionSize(1:2) 1]),Formattable=true,Name="emb_reshape")];
lgraphGenerator = addLayers(lgraphGenerator,layers);
lgraphGenerator = connectLayers(lgraphGenerator,"emb_reshape","cat/in2");
netG = dlnetwork(lgraphGenerator);
dropoutProb = 0.75;
%numFilters = 64;
scale = 0.2;
filterSize = 5;
layersDiscriminator = [
imageInputLayer(inputSize,Normalization="none")
dropoutLayer(dropoutProb)
concatenationLayer(3,2,Name="cat")
convolution2dLayer(filterSize,numFilters,Stride=2,Padding="same")
leakyReluLayer(scale)
convolution2dLayer(filterSize,2*numFilters,Stride=2,Padding="same")
batchNormalizationLayer
leakyReluLayer(scale)
convolution2dLayer(filterSize,4*numFilters,Stride=2,Padding="same")
batchNormalizationLayer
leakyReluLayer(scale)
convolution2dLayer(filterSize,8*numFilters,Stride=2,Padding="same")
batchNormalizationLayer
leakyReluLayer(scale)
convolution2dLayer(Factor*4,1)];
lgraphDiscriminator = layerGraph(layersDiscriminator);
layers = [
featureInputLayer(1)
embeddingLayer(embeddingDimension,numClasses)
fullyConnectedLayer(prod(inputSize(1:2)))
functionLayer(@(X) feature2image(X,[inputSize(1:2) 1]),Formattable=true,Name="emb_reshape")];
lgraphDiscriminator = addLayers(lgraphDiscriminator,layers);
lgraphDiscriminator = connectLayers(lgraphDiscriminator,"emb_reshape","cat/in2");
netD = dlnetwork(lgraphDiscriminator);
However, the above code gives an error at
[~,~,gradientsG,gradientsD,stateG,scoreG,scoreD] = ...
dlfeval(@modelLoss2,netG,netD,X,T,Z,flipFactor);
The size of generated image at
[XGenerated,stageG] = forward(netG,Z,T);
is 256x256x3. However, an error comes stating that trainedVariance is not positive
Could you assist me which transposedConv2dLayer to change to adjust the size to 224x224x3 or 256x256x3?
Thanks for your help

请先登录，再回答此问题。

回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Image Data Workflows 的更多信息

标签

产品


版本

R2022a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by