search

ajbrock / BigGAN-PyTorch

The author's officially unofficial PyTorch BigGAN implementation.
MIT License
2.84k stars 470 forks source link

Mode collapse? #75

Open Hierakonpolis opened 3 years ago

Hierakonpolis commented 3 years ago

Hello, I tried training on ImageNet with the default script using 4 GPUs, and the GAN stopped training generating something like this: fixed_samples1142000 All generated samples are similar grids. Am I missing something and are there any parameters I should be tweaking, or should I just try again?

datduong commented 3 years ago

I also run into similar problem but with my own dataset. I messed around with argument z_var and got just slightly better results. I am looking into WGAN to avoid this problem.

fido20160817 commented 2 years ago

same happened to me.

fido20160817 commented 2 years ago

when I trained on BigGANdeep model, things come normal, and you can try this model. But I still don't know what's the reason of the collapse of BigGAN.

xiaomao19970819 commented 2 years ago

@fido20160817 Hi buddy, can you provide a pre-training model of biggan-deep? I see this repo is not provided