Adrien Corenflos

Can you relax style requirements for notebook? I ran the pre-commit locally, it reformatted the notebook, but it's not working here. > reformatted notebooks/TemperedSMC.ipynb > All done! ✨ 🍰 ✨...

https://www.jstor.org/stable/4140600?seq=1, Section 4.2. I didn't think this was still considered a valid approach in the continuous case, but maybe I'm wrong. It's perhaps more interesting in the discrete case: https://arxiv.org/pdf/1101.6037.pdf

Sorry to jump in but that's actually easier than you seem to think. See the below code as an example ```python from functools import partial import numba as nb from...

Ah yes that sounds tricky if you don't want to break the open close principles. Also even doing it on a case by case basis may prove complicated. For example,...

Hi all, I was reading through this, and I'm surprised none of you mentioned using tensorflow as a backend at all. May I ask why it is not being considered?...

> @AdrienCorenflos Generally, from my experience, torch is a lot more intuitive to use vs. tensorflow. It feels more pythonic. This same view is held by many users, as evidence...

> my big personal concern currently is code duplication between gpu and non-gpu code. The GPU code is not maintained or tested as well as the CPU code. See how...

I personally have the following in tf2 which I can contribute without much effort (apart from interface style maybe) - geomloss (a personal translation of feydy's code) - standard sinkhorn...

I agree with the above. We have had several chats with @rlouf over the past on this specific point. **Food for thought:** The "only" way I see to do this...

An additional point to note: at the moment, at least the Gaussian sampler is parametrised in a way that reduces the need for computation: https://github.com/blackjax-devs/blackjax/blob/6b76746e0352c71c2489b12d38b8457ffb1ab13e/blackjax/mcmc/marginal_latent_gaussian.py#L46 Changing the tempering parameter from...