question about conv for mu and sigma in autoencoderkl

[Can-Zhao]

Hi,

In AutoencoderKL, the mu and sigma are results of convolution from h,

GenerativeModels/generative/networks/nets/autoencoderkl.py

Lines 701 to 702 in b1c1a45

	z_mu = self.quant_conv_mu(h)
	z_log_var = self.quant_conv_log_sigma(h)

.

If h has only one latent channel, then this single channel value needs to have the ability to generate both mu and sigma. It seems difficult intuitively?
I was wondering if you could help me with it. @ericspod Or should I set latent channel to be >1 ?

Thank you!

Best,
Can

[ericspod]

I would guess all that matters is how large h is in total, even with 1 channel it's spatial dimensions might be large enough that it has sufficient capacity to represent a compressed form of the input. I would assume this would be like any other VAE in that if your results after encoding/decoding are poor then your capacity is too small and so adding channels could help with that. @Warvito might have more insight on this.

[Can-Zhao]

Thank you so much Eric! Yes, if h is large enough, then channel=1 is also fine