optimize() keeps all UserWarnings disabled

[alexzerg]

I found this while checking optimize.py and its tests. There is a filterwarnings("ignore", "", UserWarning) call inside optimize(), but looks like the filter stays active after the function is already done. I tried this on current joss2.0 python 3.13:

import warnings
import numpy as np

from pynumdiff.optimize import optimize
from pynumdiff.smooth_finite_difference import butterdiff

x = np.sin(np.linspace(0, 2*np.pi, 40))
truth = np.cos(np.linspace(0, 2*np.pi, 40))

warnings.resetwarnings()
print(warnings.filters)

optimize(
 butterdiff,
 x,
 1.0,
 dxdt_truth=truth,
 parallel=False,
 maxiter=1,
 search_space_updates={
 "filter_order": 2,
 "cutoff_freq": [0.1],
 "num_iterations": 1,
 },
)

print(warnings.filters)
warnings.warn("test warning”)

After optimize() the filters are:

[('ignore', None, <class 'UserWarning'>, None, 0)]

and the last warning doesn't print. So one optimizer call can also hide unrelated warnings later in the same python process.
Maybe the filtering should stay local to the optimizer call and restore whatever filters the caller had before?

Review reference: openjournals/joss-reviews#11172

[pavelkomarov]

Ah, again good eye. The issue here was that warnings might get triggered in subprocesses as you're trying to optimize something, and it can be overwhelming at the level of the optimizer to see the flood of them come in. For example, certain choices of hyperparameters might cause a failure internal to a method that makes it return a bunch of NaNs, but the optimizer just throws that result out and looks for a hyperparam combo that works and returns a good result. Silencing was a bit of a challenge, so I eventually got pretty aggressive, but I didn't realize that leaves warnings turned off for future, unrelated function calls. We want the scope to be local to optimize(), certainly.

[pavelkomarov]

e8a2f92

[alexzerg]

Retested on current joss2.0, filters are restored after optimize() and warnings are visible again.
Thx!