about architecture

[pru12345]

Render task - pose VAE was used to encode the pose to Z(y), but in VAE the hidden variable is calculated by (m+sd*E) where E is a random sample from a Gaussian distribution. here is my question - you had used the path (render) for training Ali net, for the same datapoint of pose there would be the different z(y) because of random E, how the encoder (VAE) path is used for train task.

[melonwan]

Thank you for the interest. For later training, we limit the variance of the Gaussian distribution the noise is sampled from. To this end, it works similar to the variational dropout proposed by DP Kingma, who also proposed VAE.