[BUG] n_jobs validation is incorrect when it is negative

[yarnabrina]

This is the current logic:

sktime/sktime/utils/validation/__init__.py

Lines 126 to 148 in ca8d8e6

	def check_n_jobs(n_jobs: int) -> int:
	"""Check `n_jobs` parameter according to the scikit-learn convention.
	Parameters
	----------
	n_jobs : int, positive or -1
	The number of jobs for parallelization.
	Returns
	-------
	n_jobs : int
	Checked number of jobs.
	"""
	# scikit-learn convention
	# https://scikit-learn.org/stable/glossary.html#term-n-jobs
	if n_jobs is None:
	return 1
	elif not is_int(n_jobs):
	raise ValueError(f"`n_jobs` must be None or an integer, but found: {n_jobs}")
	elif n_jobs < 0:
	return os.cpu_count() - n_jobs + 1
	else:
	return n_jobs

If n_jobs is negative, are we not requesting more CPU than that is available in L145-146?

I think we may use this, probably always:

https://github.com/joblib/joblib/blob/6310841f66352bbf958cc190a973adcca611f4c7/joblib/test/test_parallel.py#L611-L614

[fkiraly]

hm, I did not know this code existed.

I thought the validation would be done anyway by joblib again.

The new parallelization abstraction entry point sktime.utils.parallel does not use it, btw.

Should we just get rid of it and let joblib handle it?

[yarnabrina]

I found this in #5989 who was modifying this function for n_jobs=-1 case, so I don't know if this is being used or not. If unused, removal makes sense.

---

It's being used I think.

[fkiraly]

I know (well, since 5 min ago 😁 ), I found the same. Seems mostly used from:

• transformers in the panel submodule
• upon cursory inspection, in repeated boilerplate that could/should be replaced by base class broadcasting and parallelization