Open DylanMannKrzisnik opened 2 years ago
Hi,
I'd say the simplest would be to start from e.g. this example: https://github.com/didriknielsen/survae_flows/blob/master/examples/mnist_vae.py and add some bijective transformations after the VAE on line 50 (the layers are specified from data to latent, i.e. x -> z).
Alternatively, and equivalently, you can replace the base_distribution
with a Flow
.
Hope this helps. If it's not clear I post an example later.
Didrik
Thanks for your response, I had opted for the second option of replacing the base_distribution with a Flow. Much appreciated.
Hello,
What would be the simplest way of replacing a Gaussian prior for VAE with a flow-based one?
Thanks