Fix #298 Handle removal of sklearn randomized l1 models

[teabolt]

Fixes #298
Wrap import and decorators in a try-except block.

[codecov-io]

Codecov Report

Merging #302 into master will decrease coverage by 0.06%.
The diff coverage is 71.42%.

@@            Coverage Diff             @@
##           master     #302      +/-   ##
==========================================
- Coverage   97.26%   97.19%   -0.07%     
==========================================
  Files          44       44              
  Lines        2851     2854       +3     
  Branches      541      541              
==========================================
+ Hits         2773     2774       +1     
- Misses         41       43       +2     
  Partials       37       37

Impacted Files	Coverage Δ
eli5/sklearn/transform.py	94.28% <71.42%> (-5.72%)	⬇️

[lopuhin]

Looks great @teabolt , thank you! Left some comment, but I think PR is good to be merged as-is.

scikit-learn 0.21 is not released yet, but it would be picked up on CI once it's on PyPI, so good that we'll have less breakage then.

[lopuhin]

SelectorMixin could be left as a decorator, right?

[teabolt]

You are correct. I expanded out and placed SelectorMixin last to keep the original order that the decorators are applied in (RandomizedLasso -> RandomizedLogisticRegression -> SelectorMixin). But in this case order does not seem to matter.

[lopuhin]

I had an idea for another approach here, but after some thought it looks worse than what is implemented here, still let me put it here:

• put imports at the top, set RandomizedLasso etc. to None if they fail to import
• make it possible to write something equivalent to @transform_feature_names.register(RandomizedLasso) which would do nothing if RandomizedLasso is None. But it seems that we'd need a wrapper for that, something like @register(transform_feature_names, RandomizedLasso)

Advantage is that the flow would be more natural (imports at the top, decorators above the function). Disadvantage is that it might be a bit confusing that RandomizedLasso can be None, and also we would need to introduce the wrapper. So to me it seems that it would make sense only if we get a lot of similar conditional registers.

[teabolt]

Thanks for the suggestions @lopuhin. Looking at the recently deprecated in sklearn there doesn't seem to be too many removals of existing models. But I would certainly prefer your approach if we face more issues like this down the line.

[lopuhin]

Looks great, thanks @teabolt