jeig 0.0.2

Various eigendecomposition implementations wrapped for jax.

jeig - Eigendecompositions wrapped for jax

v0.0.2

Overview

This package wraps eigendecompositions as provided by jax, numpy, scipy, and torch for use with jax. Depending upon your system and your versions of these packages, you may observe significant speed differences.

The wrapped eig function also includes a custom vjp rule so that gradients with respect to eigenvalues and eigenvectors can be computed.

Install

jeig can be installed via pip,

pip install jeig

This will also install torch. If you only need torch for use with jeig, then the CPU-only version is sufficient and you may wish to install manually as described in the pytorch docs.

Example usage

import jax
import jeig.eig as jeig

matrix = jax.random.normal(jax.random.PRNGKey(0), (8, 320, 320))

jeig.BACKEND_EIG = jeig.JAX
%timeit jeig.eig(matrix)

jeig.BACKEND_EIG = jeig.NUMPY
%timeit jeig.eig(matrix)

jeig.BACKEND_EIG = jeig.SCIPY
%timeit jeig.eig(matrix)

jeig.BACKEND_EIG = jeig.TORCH
%timeit jeig.eig(matrix)

376 ms ± 11.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
689 ms ± 11.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
414 ms ± 19 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
136 ms ± 4 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

Credit

The high-level eig function and the tests are adapted from fmmax. The torch implementation of eigendecomposition is due to a comment by @YouJiacheng.