search

shaanrockz / COCO-GAN

PyTorch implementation of COCO-GAN (https://hubert0527.github.io/COCO-GAN/)
MIT License
25 stars 0 forks source link

Two issues regarding the implementation #1

Open hubert0527 opened 4 years ago

hubert0527 commented 4 years ago

Thanks for your interest in COCO-GAN and your effort in implementing it in Pytorch!

I took a quick look and found two issues (not sure if there is still any other) that may cause your model failed to converge:

  1. The implementation of CBN is wrong, the mean_rec and var_rec needs to be tracked for each iteration. It should work very similarly to the standard batch norm. You may refer to this: https://discuss.pytorch.org/t/conditional-batch-normalization/14412/2

  2. You will have to implement a GPU version for micro-patch-to-macro-patch conversion. Otherwise, there will be no gradient propagation from the discriminator for the generator adversarial training.

shaanrockz commented 4 years ago

Thank you for pointing out, appreciated. Will update the code. The code was debugged and run on the cpu for implementation, I should have added this in the description, my bad. Anyways I will also add the gpu implementation of the functions soon.

waydegg commented 4 years ago

Have these issues been fixed yet?

shaanrockz commented 4 years ago

Have these issues been fixed yet?

Hi, sorry for late reply. I will update the codebase soon.