how to fulfill GAN (Generative Adversarial Networks ) or DCGAN in matlab?

I find that there is no example demo for GAN (Generative Adversarial Networks ) or DCGAN. I wonder how to fulfill GAN in matlab? if for GAN, is the last output of the generator RegressionOutputLayer or others?

[Kenta]

Please check the file below. This is an example of conditional GAN.

https://jp.mathworks.com/matlabcentral/fileexchange/74921-conditional-gan-generative-adversarial-network-with-mnist

Answer 1

[Muhammad Usama Sharaf SAAFI]

Only MATLAB 2019b has demo example of GAN.Example code also works on GPU but you should have CUDA 10.1 driver installed in your system However you can also look below link if you donot have Matlab 2019b.

https://github.com/JZhaoCH/GAN-Base-on-Matlab

Answer 2

[Walter Roberson]

Generative Adversarial Neural Networks are not available in any Mathworks product. They are not supported by the Neural Network toolbox.

[Jack Xiao]

what a pity! looking forward to the supporting!

[Jony Castagna]

Just looked at version 2019b: they support GANs now!

Answer 3

[mohammed mahmoud]

View this link dcgan-matconvnet

Answer 4

[azad]

still no GAN support !

[Jack Xiao]

Yes, so sad... hope coming soon...

Answer 5

[Niklas]

In our team we realize GANs with regression layers as output layer. Works fine in our cases.

You also have the possibility to define own layers: See this doc https://de.mathworks.com/help/deeplearning/ug/define-custom-deep-learning-layers.html for further Informations.

Best wishes, Niklas

[Jack Xiao]

thank you, i wonder how do you train discriminator as there are still no demos for training multiple nets simultaneously at present?

can you release your project or any detailed information?

[Niklas]

Unfortunately I can not upload the code.

We use a RBM for our generator model. This we pretrain with own code but this part is very similar to the code snippets from Geoffrey Hinton. We also train a CNN for the descriminator part before we stick both models together. So after pretraining the CNN is very good in decisions whether an image is generated or a real one.

Afterwards we use some tricks. First we created a custom layer for the input where we are able to insert the seed for the generator and a real image. We pass the real image around the RBM. (Please note that you need a custom Sigmoid Activation Layer for RBMs). Afterwards we use custom layers in the CNN to identify both images simultaneously with the same weights as from the pretraining. You can just copy the Matlab Layers for that and modify them a bit. Finally we created an output layer for this custom CNN. There we track the differences between both classification probabilities. Note that you have to think about a correct Loss Function there. We train the whole model by inserting real images with different random seeds and a final result of 0 difference between both classifications.

Just a note on "easyness". We only do that because we need MATLAB Code at the end to export them on our embedded systems. If you just want to discover what GANs can do, you should for now stick to TensorFlow as its much easier. They have a good example on their website. But maybe Mathworks add native support for GANs in a future release?!

Answer 6

[Jony Castagna]

Just looked at version 2019b: they support GANs now! If only the Matlab base + NN toolbox was free...

[Walter Roberson]

If only food and rent and property taxes were all free so that Mathworks employees didn't need to be paid... ?

Answer 7

[Yui Chun Leung]

I implemented different types of GANs with Matlab, including DCGAN, CycleGAN and more.

You can find my files in FileExchange or Github (https://github.com/zcemycl/Matlab-GAN).