Fabian Pedregosa

thanks for the clarification. I agree that both of these points would be nice improvements. For the first point, for example in the [stochastic optimization page](https://jaxopt.github.io/stable/stochastic.html) , I don't think...

@amir-saadat : would you be willing to submit a pull request implementing (some of) this?

I don't have a solution for this, but can confirm that it happens also with the `update` API, i.e., when updates are run inside a for loop: ``` def optimize(min):...

Some updates on my investigations. 1. Upon @mblondel's idea, I set eq=True in the definition of LBFGS. It didn't help. 2. I also modified the LBFGS class to remote the...

yeah, although there's a small decrease at the end that could mean it's recuperating some memory. This is without jitting:  and with jitting:  As you can see, it's...

Yes to both. Seems like a bug and should be fixed (although we're all spread too thin, I wouldn't know how to set a timeline on it) On Thu, Feb...

This behavior can be avoided using the newly implemented ` jax.clear_caches()` in jax (thanks @froystig !). For example, the code below doesn't have the ever increasing profile. Instead, it has...

I'm going to close the issue for now, but please reopen if problem persist (BTW you might need the development version of jax for the clear_caches() function)

Maybe, but at this point it seems more of an issue concerning jax than jaxopt, wdyt? On Fri, May 19, 2023, 11:30 Mathieu Blondel ***@***.***> wrote: > It's nice to...

@froystig made a good point in private conversation, that this might be symptomatic of jaxopt not using the cache properly and/or generating too many fresh functions instead of re-using the...