VAE without sampling observation

Misc.
1

Is sampling the observation (i.e. in the VAE tutorial) always necessary in Pyro? For example, building a vanilla VAE with conv layers for colored images (as opposed to MNIST) are you not allowed to have P(X | Z) be deterministic even if Z is sampled?

Thank you

2

if Z is the same dimension as X then you can use a bijective transformation like a normalizing flow. see e.g. here

1 Like