Need help in debugging the stylegan code

Course Q&A Generative Adversarial Networks (GANS) Build Better Generative Adversarial Networks
1

I’ve built the stylegan and its training loop using Tensorflow from its research paper, I’m using celebA-HQ dataset of 30k images. But even after running more than 5000+ steps it is not showing any improvements. Can anyone help me in my script where I’m doing mistake?
PFA colab notebook

Generated image after running more than 5000 steps:

image

2

@Janvi_Chokshi welcome to this community!
Yes, I also had some problems with this task when I had my first contact with this dataset some years ago. My suggestion is to take a look at the tutorial below where you can compare your code and hyperparameter configuration and see where you can improve your solution. Note that is not easy to get a good result.

Keep learning!

3

Thanks @carlosrl for taking time to give me a feedback. I will also try the DCGAN, however can you give me reference to which I can compare my stylegan code?

4

Yes, the Tensorflow site has great tutorials about how to use the library to train different models.
In your case, you can take a look the link below that, even using a different dataset, you still can compare the discriminator, generator, and so on, and see where you can improve your model.

Keep learning!

5

Thanks a lot @carlosrl