squarem-jaxopt 0.1.3

SQUAREM accelerator for JAXopt solvers

squarem-JAXopt

squarem-JAXopt is an implementation of the SQUAREM accelerator for solving fixed-point equations, see Du and Varadhan (2020). SQUAREM is implemented in JAX and JAXopt. The later allow for implicit differentiation of the fixed-point.

Installation

pip install squarem-jaxopt

Usage

import jax
import jax.numpy as jnp
from jax import random

from jaxopt import FixedPointIteration, AndersonAcceleration
from squarem_jaxopt import SquaremAcceleration

# Increase precision to 64 bit
jax.config.update("jax_enable_x64", True)

N = 4

a = random.uniform(random.PRNGKey(111), (N, 1))
b = random.uniform(random.PRNGKey(112), (1, 1))


def fun(x: jnp.ndarray) -> jnp.ndarray:
    y = a + x @ b
    return y


fxp_none = FixedPointIteration(fixed_point_fun=fun, verbose=True)
result_none = fxp_none.run(jnp.zeros_like(a))

fxp_anderson = AndersonAcceleration(fixed_point_fun=fun, verbose=True)
result_anderson = fxp_anderson.run(jnp.zeros_like(a))

fxp_squarem = SquaremAcceleration(fixed_point_fun=fun, verbose=True)
result_squarem = fxp_squarem.run(jnp.zeros_like(a))