Question about params in decoder

[gitlabspy]

Hi, thanks for sharing your great works !!!! It's so amazing!
I have a question about the decoder, I wonder if you can help me out a bit.
Can you please explain what these three params are for? ( bias_xs, gain and bias )

vdvae/vae.py

Lines 186 to 190 in ea35b49

	self.bias_xs = nn.ParameterList([nn.Parameter(torch.zeros(1, self.widths[res], res, res)) for res in self.resolutions if res <= H.no_bias_above])
	self.out_net = DmolNet(H)
	self.gain = nn.Parameter(torch.ones(1, H.width, 1, 1))
	self.bias = nn.Parameter(torch.zeros(1, H.width, 1, 1))
	self.final_fn = lambda x: x * self.gain + self.bias

[pabloppp]

If I may try to answer...
I understand the bias_xs as a fixed initial parámeter, if you look at how the topdown block is described in the paper, you'll see that it expects a tensor as an input, and will add the conditional information from the encoder in the left branch, but in the first decoder layer there is no previous input, so we add a constant (although trainable) input that will be the same for every image (sort of how its done in StyleGAN)

As for bias and gain, they seem to be used to scale the output, but I am not sure why they are required since the weights & biases of the last convolution could serve the same purpose 🤔

[gitlabspy]

@pabloppp Thx! Your answer did help me a lot. The shape of final scale and shift are wired, unlike normal conv/fc's params' shape...