Patrick Kidger
Patrick Kidger
Do I understand that: 1. As in #60, you're interested in logging metrics during the solve, and want to save something at every step? 2. As you're currently using a...
Can you give me some examples of the sort if thing you're trying to save, by the way? That this has come up twice now suggests that it may be...
So given the state `y` (at each timestep), then at the moment, with `SaveAt(steps=True)`, we simply save `y`. Perhaps what could be done is to allow inserting a user-specified function...
Sounds good! I've adjusted the title of this issue to make this a feature request. (Which if you ever feel like digging into the guts of Diffrax I'd be happy...
Yep! It's this big block of code. The idea will be to change from `state.ys` -- which is where we save our output -- to `state.out`. Now the bad news...
Looks like the formatting is failing. Have a look at [`CONTRIBUTING.md`](https://github.com/patrick-kidger/diffrax/blob/main/CONTRIBUTING.md). I've not really gone through most of the example yet; I'll leave a proper review of that once everything...
Hmm. This looks like a bug in core JAX -- `closed_call` looks like a new JAX primitive, which `jax.experimental.host_callback.call` hasn't been updated to be able to handle. I'm not sure...
At first glance it looks like they're just computing Hessians by composing forward-mode autodiff with reverse-mode autodiff. This is already a standard technique, and it can be done today simply...
Sounds reasonable. I'd be happy to accept a PR doing this.