patrick-kidger / patrick-kidger/diffrax

Possible to get BacksolveAdjoint backward pass solution/stats?

Open
#110 3 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

question
Dominant language
Python
Stars
2.1k
Forks
189
Avg merge
3d 18h
Merged PRs (30d)
1

Description

Thanks for this excellent library!

When using BacksolveAdjoint adjoint method, it's easy to get (and log) forward pass stats with

sol = diffrax.diffeqsolve(
    terms, solver, t_0, t_T, initial_step_size, init_y, args,
    saveat=saveat, stepsize_controller=stepsize_controller,
    adjoint=diffrax.BacksolveAdjoint()
)
stats = sol.stats
# ...
(loss, stats), grads = jax.value_and_grad(loss_fn, has_aux=True)(policy_params, init_states)

I believe the backward pass also involves solving ODE, but I did not figure out a way to get its stats. Could you suggest how to achieve this functionality? Or is it something limited by JAX API?

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start at the diffrax.diffeqsolve entry point and BacksolveAdjoint, then trace how forward-pass sol.stats is exposed through the shown jax.value_and_grad call. Determine whether backward-solve statistics can be surfaced through this API; done means a supported retrieval path is documented or the JAX limitation is clearly explained.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
30/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.