Patrick Kidger
Patrick Kidger
As in the paper [The Geometry of Nonlinear Least Squares, with applications to Sloppy Models and Optimization](https://arxiv.org/pdf/1010.1449.pdf) by Transtrum, Machta, and Sethna.
Right now these have been implemented as standalone solvers.
- [ ] Anderson acceleration - [ ] LBFGS - [ ] Affine Powell's (unconstrained) derivative free optimisers: - [ ] UOBYQA - [ ] NEWUOA On an affine solvers:...
This would allow us to use all our line searches and descents with the nonlinear CG approximate Hessian. See [Conjugate Gradient Methods with Inexact Line Search](https://www.jstor.org/stable/3689494) by Shanno.
**Reproduction:** Given the following directories: ``` one/pyproject.toml two/pyproject.toml three/pyproject.toml ``` with contents: ```toml # one/pyproject.toml [project] name = "one" version = "0.0.1" dependencies = ["two @ file://${PROJECT_ROOT}/../two"] # two/pyproject.toml [project]...
Looks like JAX used to do some "broadcasting" here, and no longer does. Bearing in mind that a PyTree may have arrays of multiple ranks, I'm not immediately sure what...