Train GANs on a custom images.

[Un1nU]

Hello, thanks for your excellent project. I want to train the network on my own images with 128x128x1 size. But I encountered some issuses.

I have change to

self.img_rows = 128
self.img_cols = 128
self.channels = 1

and write my own data_loader() function to get X_train and y_train. But I'm confuses about how to modify the network structure correctly. I tried changing it, but the result I got was a noisy image. I am a beginner. Is there anyone help me?

[VincieD]

Hi, i've took the KERAS code for cwgan-gp as a base for my own project and build upon my code. Check this project: https://github.com/VincieD/TrafficSign_GAN_Benchmark

There you will see how i changed the structure of generator and dicriminator in order to get to 64x64 px resolution. The problem of 128x128 can be the convergence. I can imagine that your G and D will not converge. But you can try to add additional layer at the end and start training. Let me know if you need some additional hints :)

[Un1nU]

Hi, i've took the KERAS code for cwgan-gp as a base for my own project and build upon my code. Check this project: https://github.com/VincieD/TrafficSign_GAN_Benchmark

There you will see how i changed the structure of generator and dicriminator in order to get to 64x64 px resolution. The problem of 128x128 can be the convergence. I can imagine that your G and D will not converge. But you can try to add additional layer at the end and start training. Let me know if you need some additional hints :)

Thanks for your help. Your project is good. I have been changed it before. But now, I have some diffculties that adding extra conditions besides the image class (I am not sure if I express clearly). Do you have any good ideas? Thank you.