[MRG] FEA Add sequential feature selection transformer

[NicolasHug]

This PR adds a SequentialFeatureSelector transformer for forward and backward sequential selection.

Reboots and Closes #8684 @rasbt
Closes #6545

• Docs and User Guide
• Example
• Tests

To minimize reviewing efforts, I made this a minimal working version. In particular there is no support for setting n_features_to_select into a range. This should be reasonably easy to implement but I want to leave that for subsequent PRs.

[NicolasHug]

TLDR: Are we OK to change the linux 32bits CI from using atlas to using blas + lapack?

---

check_estimators_dtypes() had been failing for a while with the new SequentialFeatureSelector, on the linux 32bits CI. While I don't fully understand the issue, it's related to the call to svd in LinearDiscriminantAnalysis which eventually leads to:

On entry to DGESDD parameter number 10 had an illegal value

Switching from atlas to blas + lapack has solved the issue. It required changing the tolerance of a recently introduced (and unrelated) check in RBM.

CC @ogrisel @thomasjpfan @jeremiedbb @rth

[thomasjpfan]

check_estimators_dtypes() had been failing for a while with the new SequentialFeatureSelector, on the linux 32bits CI. While I don't fully understand the issue, it's related to the call to svd in LinearDiscriminantAnalysis which eventually leads to:

Can we extract out the dataset that causes LinearDiscriminantAnalysis to segfault?

[NicolasHug]

Can we extract out the dataset that causes LinearDiscriminantAnalysis to segfault?

IDK, can we SSH into the azure instances?

[thomasjpfan]

I do not know how much we want to support support atlas on 32 bit. I think we would need some historical context to answer that question.

Edit: The "hack" would be to change the seed in

scikit-learn/sklearn/utils/estimator_checks.py

Line 565 in c3effe4

bias=5.0, noise=20, random_state=42,

and see what happens.

[NicolasHug]

Edit: The "hack" would be to change the seed

were you able to reproduce @thomasjpfan, or is this a wild guess?

[thomasjpfan]

Wild guess.

[thomasjpfan]

I ran

python -m pytest --showlocals --durations=20 --junitxml=test-data.xml \
--pyargs sklearn.tests.test_common -k check_estimators_dtypes -v

with altas here and it passed. This is a very strange bug.

[NicolasHug]

Is there any difference between our CI and the instance you ran @thomasjpfan ? (apart from the -k part)?

This is a very strange bug.

Indeed... not to mention that the 32bits thing started failing after a commit that was just merging with master.

[NicolasHug]

? @thomasjpfan

[thomasjpfan]

32 bit is not failing anymore with the atlas.

OSX just timeouted. (This has been happening from time to time)

[thomasjpfan]

Just reran the osx test and it did not hang. All the tests pass now.

[NicolasHug]

what happened? The package version didn't even change, did it?

[NicolasHug]

Looks like the atlas issue was randomly fixed...

I addressed all you comments @glemaitre, ready to merge?

[thomasjpfan]

Looks like the atlas issue was randomly fixed...

I wasn't able to reproduce the bug when I was debugging it the other day. The "biggest" change was that pytest-xdist is not installed anymore on the 32bit instance.

[glemaitre]

Only made a numpydoc change. Thanks @rasbt and @NicolasHug