forward_kld with weighted training samples about normalizing-flows

[VincentStimper]

Hi @alexmiltenberger,

nice to hear that you enjoy the package.
Instead of modifying the method, you could also just compute log_q using this method and then applying the weights when computing your loss in an extra step. This might be more efficient since in an importance sampling context the weights might depend on log_q as well, so you can directly reuse the results and do not have to iterate through the flow twice by calling the forward_kld and the log_prob method.
What do you think?

Best regards,
Vincent

[alexmiltenberger]

Ah, didn't realize the log_prob method was almost identical to forward_kld. This is a much better solution than modifying forward_kld. I just tried it and it works well. Thanks so much for getting back to me!