Daniel Gedon comments

Results 6 comments of


                                            Daniel Gedon

dot_product = s_loss_old - s_loss_new but s_loss_new - s_loss_old?

If I am not mistaken, then the first order Taylor expansion goes as `f(x) = f(a) + f'(a)(x-a)`. So there is f'(a) instead of f'(x). Then with the same notation...

dot_product = s_loss_old - s_loss_new but s_loss_new - s_loss_old?

Thanks @hyhieu. # **About Taylor expansion:** It works out nicely when you start in your way. However, I have two follow-up remarks/question on it: 1. With your derivation I think...

Add `CategoricalMADE`

Thanks @janfb! > I dig into into the nflows MADE code to find out why this happens This assumes that the nflows MADE implementation does not work for d>1 dimensions,...

Discrete (mixed) density estimators

fixed by https://github.com/sbi-dev/sbi/pull/1362 Potentially one could open a new issue for mixed in the sense of some dimensions are estimated with one network and other with another network.

TabPFNRegressor preprocessing fails on bigger datasets

A simple fix could be to replace the lines below with `n_quantiles=min(max(num_examples // 10, 2), 10_000)`. The quantiles are then estimated from a subsampled 10_000 samples which might lead to...

Add NPE-PFN

I like the idea a lot! Thanks for checking that julius!