-
Notifications
You must be signed in to change notification settings - Fork 111
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
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
All reactions
Replies: 1 comment 1 reply
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.
All reactions
Thank you so much Eric! Yes, if h is large enough, then channel=1 is also fine