Carlos Martin
Carlos Martin
@rdyro Does this look good?
What does @vroulet think?
> any norm of an empty matrix is zero Carl de Boor. *An empty exercise*. ACM Signum Newsletter 25 (4), 2-6, 1990.
As previously pointed out, a simple reversal breaks stability, and negating fails for non-negatable dtypes. This can even lead to subtle bugs, as pointed out by [tvladyslav](https://github.com/numpy/numpy/issues/14728#issuecomment-2450063089). Since the most...
Stewart (1980) and Anderson et al. (1987) use a different technique. Mezzadri (2007) is this reference. Higham (2002) and Edelman and Rao (2005) seem to be more general references on...
@mhvk Good point. That would be more general. We could also allow multiple axes to be excluded.
@aymuos15 To see how Optax currently implements its [projection](https://optax.readthedocs.io/en/latest/api/projections.html) functions, see the [projections](https://github.com/google-deepmind/optax/blob/main/optax/projections/_projections.py) module. To see how JAXopt implements the ones in the above list, you can click on one...
I suggest creating a separate PR on top of the current head and then merging after https://github.com/google-deepmind/optax/pull/1351 is merged (or waiting until https://github.com/google-deepmind/optax/pull/1351 is merged and then creating a PR...
Below is a demonstration: ```python3 import argparse import jax import optax from jax import lax, numpy as jnp from matplotlib import pyplot as plt, rcParams def optimistic_sgd(learning_rate, strength): return optax.scale_by_optimistic_gradient(-learning_rate,...