Infinite loss with ADVI

I added some features to my model and now when ADVI is working, it shows in the progress bar “Average loss = inf”

But the elbo seems to converge and the average loss is 1.3835e+05 at the end:

Is there something I didn’t specify correctly in my model or maybe my new features causes problems?


Looking at the ELBO plot it seems fine. Maybe in some iterations the ELBO goes infinite, you can check where the approx.hist is infinite, and track the parameter using to investigate further.

1 Like