Not Implementing IWAE?

[GloryyrolG]

Hi,

Thanks for your good work and repo. May I ask did you implement the exact IWAE as a tighter lower bound, in case I missed it?

See the reference:

Likelihood-Regret/train_pixel.py

Line 130 in 5517c9b

[z,mu,logvar] = netE(x)

[XavierXiao]

We don't train the VAE by IWAE objective, we simply use the plain ELBO (or IWAE with k=1). We only compute the likelihood estimate by IWAE during test time.

[GloryyrolG]

I see. Btw, should not we only max along the dim of samples z?

Likelihood-Regret/compute_LR.py

Line 64 in 5517c9b

NLL_loss = -(torch.log(torch.mean(torch.exp(weights - weights.max())))+weights.max())