Patrick Kidger

This looks expected. Each `StateIndex` has different static values to every other state index. And thus a different PyTree structure. In practice the de/serialisation routines do understand `StateIndex` and are...

I'd leave it private for now; it's not important enough to be worth documenting IMO. Hmm yeah, so this looks like a tricky issue. Really the whole point of the...

Sorry, I reversed your original example. I think your original comment was about deserialising everything except the final linear layer? (My code was about deserialising just the final linear layer.)...

Yep, I could see some documentation being a good idea. (Including a version of my above example?)

FWIW I tend to use the PyTorch dataloaders, mostly just out of familiarity. I know many folks use TF as well. (The custom dataloaders in the docs are just to...

Yep pretty much.

That I didn't know! Thanks, that's really useful information.

Diffrax doesn't support complexes at the moment. This was a concious choice on my part, so that other aspects of Diffrax could be prioritised. My expectation is that there's probably...

I haven't looked into this so I don't know how many places might need changing. If you're happy to dig into this I'd still be happy to accept a PR...

Right now Diffrax doesn't support complex dtypes. I'd quite like for them to work, but realistically it's low down my own priority list. I would be very happy to accept...