Leerw / torchimize

optimization algorithms using pytorch

torchimize

torchimize contains implementations of the Gauss-Newton and Levenberg-Marquardt optimization algorithms using the PyTorch library. The main motivation for this project is to enable convex optimization on GPUs based on the torch.Tensor class, which (as of April 2022) is widely used in the deep learning field.

Functional API Usage

# gauss-newton
from torchimize.functions import lsq_gna
coeffs_gna, eps_gna = lsq_gna(initials, cost_fun, args=(other_inputs,), tol=1e-6)

# levenberg-marquardt
from torchimize.functions import lsq_lma
coeffs_lma, eps_lma = lsq_lma(initials, cost_fun, args=(other_inputs,), tol=1e-6)