How is this related to #9587 ? Will this change not cause that regression again?

#9604 has a regression test, right? And that passes. It's related to #9587 in that it undoes the API violation from #9604 ;)

LGTM

@jorisvandenbossche this will be a behaviour regression for users who previously did

p = Pipeline((('a', Something()),))
p.steps[0] = ('a', SomeOtherThing())

But we never really said we'd support tuples and that's a fairly obscure usage pattern.

@agramfort, I see you've taken to the "approve" button. Should that be understood as LGTM?

Ah, missed that you do the conversion to list inside fit. So that indeed passes the test I added (although the test should actually be changed slightly to re-initialize pipe in test_pipeline_init_tuple for the set_params test, to make sure set_params works before the pipe is fitted).
Note that the API violation already existed before #9604 :-) (depending on the type of the steps arg)

Also note that there is such a modification in the __init__ of FeatureUnion as well.

Well before #9604 there was no deprecation warning in the tests. I'm not sure in which case a copy or conversion happened. I'm not sure why FeatureUnion doesn't raise a warning :-/

Actually, it does...

from sklearn.pipeline import make_union, FeatureUnion
from sklearn.preprocessing import StandardScaler
from sklearn.base import clone
clone(make_union(StandardScaler()))

base.py:115: DeprecationWarning: Estimator FeatureUnion modifies parameters in init. This behavior is deprecated as of 0.18 and support for this behavior will be removed in 0.20.

@agramfort, I see you've taken to the "approve" button. Should that be understood as LGTM?

yes indeed :)

I am "old" but I can adapt to new features. Just don't make me learn another programming language that's all :)