Rémi Louf
Rémi Louf
That's what I thought. Shall we close this issue then?
It [appears](https://github.com/pytoolz/toolz/issues/188) that `toolz` does not use trampoline evaluation, and has no intention to do so. We may need to implement a version of `interleave` that uses it to avoid...
Ah. I tried in a fresh environment and the error still shows on my machine.
I'll take this up.
@junpenglao would appreciate your opinion on `window_adaptation` before I move on to refactor `pathfinder_adaptation` as well
> High level question: are we going to do away the `kernel_factory` in other adaptation scheme as well? IIUC they suffer the same hard to vmap problem. Yes! I'm perfecting...
> SG - what about other places with `kernel_factory` signature (`*tempered_smc` in blackjax/kernel.py)? Oh yes these are going to, they've been bothering me for a while now. The new warmup...
The main pain point in terms of API seemed to be adaptation with multiple chains. My original incline was to not return the adapted mcmc kernel but the values of...
I'm not sure I understand; what are you suggesting we should do? For lower-lewel usage I would personally recommend the following: ```python import blackjax schedule = blackjax.adaptation.window_adaptation.schedule(num_steps) adapt_init, adapt_update, _...
~~While I fully agree with we should adopt the `adapt_state, adapt_info` like anywhere else in the library~~ (that's a good pattern), I don't think we should use a `schedule_update` function....