questions about the prediction sub-network

[yaoqi-zd]

Hi, may I ask what's the advantages of the prediction sub-network over cascade convolutional layers in estimating the mean and std?

In your paper, you said the hyper latent representations are densely sampled to complete the estimation, without the potential information loss during convolutional operations. I'm a little confused, the prediction sub-network also contains convolutional operations, so it can also cause information loss?

btw, how much performance improvements can this prediction sub-network bring?

[huzi96]

Earlier experimental results show that the subnetwork does much better in generating the mean and scale for the entropy model. This is probably because we conduct a dense sampling of the latent representation. However, more experimental and theoretical analysis is needed to fully understand what brings in the improvement.

[yaoqi-zd]

get it, thanks for your reply!