Try AE,VAE,DCGAM,AEWGAN, VAEDCGAN on brain dataset in Pytorch
1.Dataset
2.Model
3.Results
4.Tricks for training
a.In the Discriminator or Generater/Decoder, uses sigmoid instead of tanh
b.When training the Discriminator, it would be better to freeze Generator weights update and vice verse.
e.g.
def free_params(module: nn.Module):
for p in module.parameters():
p.requires_grad = True
def frozen_params(module: nn.Module): for p in module.parameters(): p.requires_grad = False
frozen_params(G) free_params(D)
c. The network should be designed deliberately. If too shallow, the image is blurred. If too deep the image is almost black and couldn't see the brain.
- Summuray
VAE_DCGAN could help to denoise and the recovered images are also deblured compared to VAE results.