Thanks!

The install.cmd was adjusted to match the pip installed packages in building the wheels.

OK, but don't we still want to run tests with conda dependencies on Windows? What's the point of reproducing the CI for building wheels, if that is going to be tested when building wheels anyway?

I had to get the CI to use wheels to reproduce the error here. I’ll revert the back CI back to conda.

This should be noted in what's new for users of recent scipy with scikit-learn 0.21

@jnothman Whats new in for 0.22.0 or 0.21.4?

@adrinjalali Can this be in the RC to get the windows wheel to build?

What is the reason for this change?

@ogrisel Here is the PR that made the change: scipy/scipy#10067

@adrinjalali Can this be in the RC to get the windows wheel to build?

I'll include it in RC3.

Interesting, then why is only windows which fails the test? It seems like here we're reverting the behavior to something which is not the desired solution, to pass the test. According the PR in scipy, the new bounds are a better choice. Therefore we should be using them, shouldn't we?

According the PR in scipy, the new bounds are a better choice. Therefore we should be using them, shouldn't we?

In principle this would be good to do in another PR. We would need to update _nipals_twoblocks_inner_loop to take into account the new behavior of pinv2.

Interesting, then why is only windows which fails the test?

This test has a history of failing on windows:

• [MRG] Fixes test_scale_and_stability by clipping small values #13903 - 32 bit windows
• [MRG] Fixes test_scale_and_stability by reducing the decimal #13930 - randomly failing

Related:

• [MRG] Fix ARDRegression accuracy issue with scipy 1.3.0 (Issue #14055) #14067 - We vendored the scipy<1.2 version of pinvh because of the condition. There was no way to set the condition in that situation without calculate the singular values before hand.

Added to whats_new to 0.22.

I suppose we could consider this for a 0.21.4... We want to release 0.21 for CPython 3.8 and maybe it's easiest if it also supports scipy 1.3.

Thanks @thomasjpfan, I merged with your targeted whats new at 0.22. @adrinjalali please add this to your to-backport list for 0.22.X and optionally 0.21.X if you feel like making a companion 0.21.4 release to add support for Python 3.8 as suggested by @jnothman.