Fix labelprop underflow errors. #12478
Conversation
VisibleDeprecationWarning from Numpy is being ignored already and numpy seems to be moving to FutureWarning as recommended in PEP scikit-learn#565.
Also add a test.
|
It looks like the build failures are not strictly related to my changes. |
|
What warnings are you silencing? |
| warnings.simplefilter("always") | ||
| # Trigger a warning. | ||
| result = func(*args, **kw) | ||
| w = [e for e in w |
There was a problem hiding this comment.
No, please don't change this. Rather use pytest.warns which gives some precision not facilitated here.
There was a problem hiding this comment.
Okay, I'll try to rewrite the logic using pytest.warns. Thanks!
|
Ah, I see that you meant which
This fixes the |
Fixes #9313.
Supersedes #9384.
Instead of depending on the MNIST dataset in the original test, I've created a similar (random) dataset which also produced the same warnings.
Sidenote: I've had to disable
FutureWarningin addition tonp.VisibleDeprecationWarninginassert_no_warningsinsklearn.utils.testing. Perhaps this is overreaching a bit, so I could implement a new testing routine which only looks for absence of certain warnings in the output (as a separate PR?)