Anh Tong

I wonder if you want to make a tutorial example for this type of solver. The implemetation of hypersolver in `diffrax` is quite straightforward ```python class HyperEuler(diffrax.Euler): hypernet: diffrax.AbstractTerm def...

Got it! `AbstractWrappedSolver` is the way to go. I will gladly make a PR on this.

Thanks for the detailed review. I will get back on this after a few days :)

Sorry for taking so long. In the recent commits, I have changed `diffrax.misc.sde_kl_divergence` where the vector field is contructed based on the control term of the input SDE as `MultiTerm`....

@patrick-kidger Sorry for taking so long. I will try to get back to this pull request in a couple of days. What I can think of now is to handle...

Hi @harrisonzhu508, you can run the current code with `jax=0.3.15` (while waiting for further update). The latest version of `equinox=0.7.1` works fine. As Patrick mentioned, it must be something to...

Hi @patrick-kidger , I tried to make `sde_kl_divergence` can handle more general cases but I do not have a complete solution. If I understand correctly, the goal of `materialise_vf` to...

> Hi @anh-tong, thanks a lot for the very clean implementation again! I was trying to reproduce an example that is very similar to https://github.com/google-research/torchsde/blob/master/examples/latent_sde.py. Running the latter script yields...

The first-term resembles Gauss-Newton Hessian with square loss for regression problems. It can be seen at "empirical Fisher" but it may not recommended in practice (see [this paper](https://proceedings.neurips.cc/paper/2019/file/46a558d97954d0692411c861cf78ef79-Paper.pdf)). As far...

Hi, I work on something similar. For your example, what I can think of is to create two optimizers. Each optimizer transforms and update differently with different learning rates. This...