Dave Moore

I believe the model Brian posted is not exactly equivalent to ARMA, although they're quite similar. In ARMA, the autoregressive term is in terms of the past k *observed* values...

Thanks for the pointer, Jake! I'd known that a construction existed but hadn't realized how simple the change actually is. :-) We've also been looking at some applications recently that...

Thanks for the PR! I just left a few comments, but overall it looks great. I'll look forward to the updated test if/when you get a chance. (I think we...

Distribution batch slicing (e.g., `norm[:1, :]`) is close to what you want, but doesn't quite work because there's there's no way to specify repeated indices AFAIK. But I think you...

Those are great questions. The STS fitting API hides its trainable variables behind the parameters of the variational distributions, where they're not easy to access directly (this also makes it...

I don't know of recent changes that would be directly relevant to this issue. If others have found better solutions, feel free to comment!

Oh gosh, I did not expect this to become a PR! Sorry for the confusion. I've been integrating TFP JointDistributions (tl;dr: these are just probabilistic models that fit into the...

I think there are a few issues coming up here. The immediate problem IMHO is that `LogNormal(...).prob(0)` returns `nan` rather than `0` (or equivalently that `log_prob(0)` returns `nan` rather than...

I tried @happytk's solution and it does in fact work for me. @razinm is it possible something went wrong so you might not have actually been testing the modified code?...

Elbow doesn't currently support conjugate coordinate-ascent updates, which are what is often meant by "variational Bayes". But it is possible to implement LDA, and even handle large datasets using inference...