Releases: evanatyourservice/psgd_jax
Releases · evanatyourservice/psgd_jax
psgd-jax 0.2.4
What's Changed
- Swap
max_skew_triangular
formemory_save_mode
to give easy ways to use different preconditioner setups and save memory/compute. - readme updates
psgd-jax 0.2.3
What's Changed
- remove use of opt_einsum
psgd-jax 0.2.2
What's Changed
- Simplify code a little bit
psgd-jax 0.2.1
What's Changed
- add min ndim triangular arg to better catch bias and scale params for diag preconditioners without catching linear params like (512, 1)
psgd-jax 0.2.0
What's Changed
- Use jax instead of optax for init momentum
psgd-jax 0.1.10
What's Changed
- no bias correction on momentum
psgd-jax 0.1.9
What's Changed
- Precond init scale 1.0
psgd-jax 0.1.7
What's Changed
- Added new Kron optimizer, a second-order Kronecker-factored optimizer that's versatile and easy to use.
psgd-jax 0.1.5
What's Changed
- Get rid of unneeded element-wise clipping
- readability improvement
psgd-jax 0.1.4
What's Changed
- add option to clip updates elementwise to within -1 and 1
- rename
gradient_clip
toupdate_global_norm_clip
because optimizer outputs are clipped, not inputs