About the first stage and second stage

[clearlyzerolxd]

Hello, I have some questions about the way to normalize data. Should normalization be done this way to get good results or is it like this?

T.ToTensor(),

T.Normalize(mean=0.5, std=0.5)

or

T.ToTensor()
data = (data -1)/2

[clearlyzerolxd]

[clearlyzerolxd]

The reconstruction effect is still a bit blurry. Is this trend correct?

[clearlyzerolxd]

I rewrote the second stage code using pytorch according to the author's ideas. I found that it didn't seem to converge well.

[clearlyzerolxd]

The idea is roughly like this

            with torch.no_grad():
                out = vae.module.encode(data.to(device))
            loss = diffusion(out,None)
            net_optim.zero_grad()
            loss.backward()
            net_optim.step()