Question on how to reconstruct images

[MrHuff]

Hi!

Thanks a lot for this repo, it's really great! I am wondering how to reconstruct images, since the current reverse method seems to take a list of images with different sizes.

I tried writing a method:

    def reverse_data(self,z):
        for i, block in enumerate(self.blocks[::-1]):
            if i == 0:
                eps = torch.randn_like(z)
                input = block.reverse(z,eps)

            else:
                eps = torch.zeros_like(input)
                input = block.reverse(input,eps)

        return input

But I am unsure whether this is correct. I would appreciate any help!

Thank you!

[rosinality]

First you need to gather zs from blocks. Then you can reverse using these zs, but you will need to skip sampling (reparameterization) using prior. You can may refer to this: 4beed5d.

[MrHuff]

Hi I just tried it out and it works perfectly now! Thank you very much your help!

[koriavinash1]

for reconstruction do we need to normalize zs (obtained during the forward pass)?

Currently, I'm training glow on mnist with affine coupling, during the forward pass zs blows up to 100k sometimes, due to which reconstructions are Nan always. Let me if normalization is required or if I'm missing any other step during reconstruction.

Thank you!

[rosinality]

@koriavinash1 No, you don't need a normalization. It should be enough to give zs from forward to reverse with reconstruct=True.

[koriavinash1]

@rosinality thanks for the quick response, do you have any idea on how to resolve the NaN issue?

[rosinality]

@koriavinash1 You can try to reduce --n_flow and --n_block to make the model more stable.