Geoffrey Negiar

Complex gradients

4

Opening this as a nota bene. When optimizing over complex parameters, the gradient must be conjugated. Currently, all jaxopt optimizers would be incorrect on complex parameters, due to this. Moreover,...

Added Pseudo-inverse preconditioner for EqQP.

1

This allows to precompute a preconditioner, and share it across multiple outer loops, where the inner loop is solving an Equality Constrained QP. This should provide speedups when the parameters...

Linear solver benchmarking on EqQPs

5

I benchmarked GMRES (currently default in jax.eq_qp) vs other `scipy` solvers: [minres](https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.linalg.minres.html#scipy.sparse.linalg.minres) and [LGMRES](https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.linalg.lgmres.html#scipy.sparse.linalg.lgmres). I sampled random equality constrained QP KKT matrices and targets, and found pretty stark differences between...

Hi, Thanks for the cool package! I've just started playing around with it on a QP I need to differentiate through. I'm getting the following error when calling `jax.jacobian` on...

I'm trying to normalize a string using your parser, and replace all dates with a "" token. I can't find an easy way to do so, any suggestions?

Operations should allow shape broadcasting

5

It seems like the current implementation doesn't allow broadcasting arguments. Here's an example for normalizing leafs. ```python import tree_math as tm import jax import jax.numpy as jnp a = jnp.ones(10)...

Added sublinear stepsize to FW overlap example

1

Seems interesting that the sublinear step size never chooses the same two vertices in a row. My guess is that this is due to zigzagging behavior. 

Parallelization of fast_csr_mv

6

Parallelization for the sparse matrix multiplication, row-wise. Cf [this thread](https://stackoverflow.com/questions/46924092/how-to-parallelize-this-python-for-loop-when-using-numba), which announces nice speed ups.

Benchmarking algorithms of various families

7

It would be nice to have an example to compare speed of convergence of SAGA/SVRG/SFW on problems attaining the same optimum.