mpnewton 0.0.2

A package implementing mixed-precision optimizers and tools to analyze their convergence.

mpNewton

mpNewton is a python library that leverages NumPy/CuPy to provide a set of mixed precision optimizers for unconstrained optimization problems, as well as interfaces to implement custom ones. An implementation of precision-aware Conjugate Gradient is also available.

Get started with mpNewton

Run the following command to install the mpNewton package from PyPI:

pip install mpnewton

Run the following commands to install from conda:

conda config --add channels conda-forge
conda config --add channels mpnewton
conda install mpnewton

First steps

Optimize a given loss using mixed precision Newton:

import numpy as np
from mpnewton.optimizers.newton import Newton
from mpnewton.losses.examples.sinreg import SINREG
from mpnewton.losses.data_generator import DataGenerator
from mpnewton.optimizers.precision import PrecisionConfig

data_generator = DataGenerator(m=100, w_star=[1.0, 2.0])
sinreg, _ = SINREG.from_generator(generator=data_generator)

prec_set = PrecisionConfig(
    grad_dtype=np.float128,
    update_dtype=np.float64,
    solver_dtype=np.float32
)
optimizer = Newton(precision=prec_set)

w_star = optimizer.optimize(sinreg).w

Authors

• Giuseppe Carrino
• Elisa Riccietti
• Theo Mary
• Nicolas Brisebarre

Contribute to mpNewton

You can contribute to mpNewton with bug report, feature request and merge request.

Useful links

• Link to the documentation: Gitlab pages

Running unit tests

    python3 -m unittest discover -s tests -p "test_*.py"