Resolve exploding gradient in VAE

Question

How do you resolve exploding gradient in a deep generative model(VAE)?

NB: the data-set contains a lot of NaNs values in the columns

Answer 1

I'm not sure exploding gradients are typical for a VAE. This might be a different problem in your code. However, exploding gradients are typically solved via gradient clipping:

weights = weights*clipping_threshold/l2_norm(weights)

Where clipping_threshold is equal to 1.

Answer 2

VAEs can't cope with missing data by default. Clean your data first or try to apply some method for dealing with missing data i.e.:

https://arxiv.org/abs/2006.05301