Repository navigation
Adjusting tolerance for solvers #35
Description
Hi, Thanks for writing this useful package.
It looks like FISTA isn't respecting values passed to tol. I haven't checked other solvers or poked around much otherwise. See below to reproduce.
from yaglm.Glm import Glm
from yaglm.config.penalty import Lasso
from yaglm.solver.FISTA import FISTA
from yaglm.toy_data import sample_sparse_lin_reg
X, y, _ = sample_sparse_lin_reg(n_samples=100, n_features=1000)
mod = Glm(loss='lin_reg', penalty=Lasso(), solver=FISTA())
# ~~~~~~~ <- 'lasso' in Readme.md, but isn't supported
mod.fit(X, y)
# /path/to/lib/python3.10/site-packages/scipy/sparse/linalg/_eigen/_svds.py:314: UserWarning: Exited at iteration 20 with accuracies
# [0.93402857]
# not reaching the requested tolerance 3.162277660168379e-06.
# eigvals, eigvec = lobpcg(XH_X, X, tol=tol ** 2, maxiter=maxiter,
# etc.
import numpy as np
mod = Glm(loss='lin_reg', penalty=Lasso(), solver=FISTA(tol=np.inf))
mod.fit(X, y)
# same UserWarning
print(mod.solver.tol)
# inf
Activity