Ch06. The problem that Efficient GAN learning is not done properly

[gymoon10]

Hi, I am a student studying DL with your book. Thank you for providing high-quality learning contents and programming code.

I'm asking the question because there was a problem of GAN training that trained GAN(Colab, with 1500 epochs) do not generate plausible images.

In fact, I observed that the Epoch_D_Loss, Epoch_G_Loss, Epoch_E_Loss continued to be 0 after a particular epoch. (https://github.com/gymoon10/Computer-Vision-Practice/tree/main/pytorch_advanced/Efficient_GAN_Problem)

I'm studying using a Colab and I wonder if it's a problem of package or version. I'm sorry for asking with English as I'm not good at Japanese. I'm sorry to trouble you, but I'd appreciate your cooperation.

[YutaroOgawa]

@gymoon10 -san

Thank you for reading my book of maybe Korean Edition.

Since the version of Pytorch is newer than this book, we will need to make some changes according to issue 144.

#144

Just now, I've prepared a google colab version.

https://github.com/YutaroOgawa/pytorch_advanced/blob/master/6_gan_anomaly_detection/6_4_EfficientGAN_GoogleClab.ipynb

I hope this notebook will help you.

Best Regards,
Yutaro

[gymoon10]

I confirmed it works well. Thank you so much for your help.

[YutaroOgawa]

@gymoon10 -san

I'm grad to hear that you say.