[REVIEW] Multiclass meta estimator wrappers and multiclass SVC

[tfeher]

This PR closes #959.

This PR implements multi class classification meta estimators, and multi class SVC.

The multiclass meta estimators are just thin wrappers around the scikit-learn counterpart to provide array conversion.
This has some overhead, but should be small compared to SVM training time. More info in issue #2876.

Limitations compared to scikit-learn multiclass SVC:

• Currently we do not handle sample_weight in multi class case, and a warning is issued. To handle that we would need to fork the multi class meta estimators and ensure that the class weight parameter is propagated to the binary classifier.
• decision_function_shape arg not supported
• break_ties arg not supported

Using sklearn.multiclass allows us to choose between one-vs-one and one-vs-rest classifiers. I have introduced an extra arg multiclass_strategy to let the user choose. One vs one is more accurate, but more costly.

This is marked as work in progress, because the actual way of mapping back and fort to numpy data types (in order to make use of sklearn.multiclass) is going be simplified by #3040. Ideally we should merge this after #3040, but if that gets blocked we can bring this forward.

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[dantegd]

Have one question regarding object consumption

[tfeher]

I have improved the docstring. Additionally I have remove the n_jobs parameter because that might mislead people to think that these wrappers can do multi GPU, which they cannot.

[codecov-io]

Codecov Report

Merging #3092 (d57fa0b) into branch-0.18 (2e4388d) will increase coverage by 0.10%.
The diff coverage is 89.28%.

@@               Coverage Diff               @@
##           branch-0.18    #3092      +/-   ##
===============================================
+ Coverage        71.45%   71.56%   +0.10%     
===============================================
  Files              205      207       +2     
  Lines            16594    16702     +108     
===============================================
+ Hits             11858    11952      +94     
- Misses            4736     4750      +14     

Impacted Files	Coverage Δ
python/cuml/multiclass/multiclass.py	84.21% <84.21%> (ø)
python/cuml/svm/svm_base.pyx	94.27% <91.30%> (-0.63%)	⬇️
python/cuml/svm/svc.pyx	95.16% <92.98%> (-0.62%)	⬇️
python/cuml/multiclass/__init__.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2e4388d...d57fa0b. Read the comment docs.

[tfeher]

Fixed failing tests:

• The estimator arg has no default param therefore the unfit_pickle and unfit_clone tests ar set to xfail.
• The docsting tests did not pass, I believe the parameter docs are not parsed correctly. For now I have skipped these test. @mdemoret-nv could you have a look while these fail?

[tfeher]

rerun tests

[tfeher]

rerun tests

[tfeher]

rerun tests

There seem to be on intermittent error with FilTests/PredictSparse16FilTest.Predict/17 on centos.

[mdemoret-nv]

Have a couple of suggestions regarding estimator usage, documentation, and decorators but overall it's nothing major.

I think before this gets merged you need to add the classes to the API doc so these new estimators will show up in our documentation.

[tfeher]

Thanks at @mdemoret-nv for the review, I appreciate your suggestions, very useful! I have addressed the issues.

I have added a link to the API doc (f36f2d5), probably @JohnZed should have a look if that is the right place.

[tfeher]

Thanks again @mdemoret-nv for the comments, I have fixed the remaining issues.

[mdemoret-nv]

Updates look good. Changes LGTM