Hi all! Are there any news on this front (either with ADVI or other methods)?
I think I’m confronted to a similar problem with a multinomial softmax regression: The model samples smoothly - no divergences and chains look good - although very slowly (2 hours and a half for 8000 samples on my machine). Priors have been carefully checked with prior predictive checks.
I first thought that sampling was hurt by some of the categories spending time near zero, but after some tests, I think the biggest burden on sampling speed is the amount of data (not that big, but high dimensions: 1889 rows for 7 categories).
So my question is two-fold:
- Although the chains look good and there are no divergences, is there still a way to optimize the model for sampling?
- Or is it just MCMC having a hard time sampling in high dimensions? In that case, which solution do you recommend? Running on a virtual machine?
Thank you very much for your help