Basic multiprocess in AutoML loop, Logger and many more quality of life changes

[vincent-antaki]

Another PR with a many quality of life improvements!

Multiprocess

This PR introduces a very simple multiprocess interface for AutoML loop. The AutoML loops now takes a "multiprocess" boolean argument and a "n_processes" integer argument in input. If multiprocess is False then the AutoML loop is executed in a single process, if it is True, then a multiprocess Pool is instanciated with n_processes process and provides a trial for each process to perform until all trials are complete.

This doesn't fulfill #297, because trial splits aren't parallelized. But I think it would be a simple modifications to also get this behaviour down to the trial split level.

Logger

Instead of prints, we now have use python's logging library! In an AutoML loop, we use a different file-based logger for every trial which are saved in the HyperparametersRepository cache folder. This closes issue #391. (And closes #255, closes #257, closes #283 and closes #298 which are the basically the same issues)

Continue-break-statement-for-each-data-input

I've retrieved Alex's last commit of his Continue-break-statement-for-each-data-input branch (PR #319) and applied the review comments. Also introduce ExecuteIf class which allows us to easily introduce logical if in our pipeline design. These add-ons partially full fill issue #443 by adding BreakIf, ContinueIf and ExecuteIf.

Introduction of a DeprecatedMetaClass.

Created a utils.py file and added a DeprecatedMetaClass which allows us to rename classes without breaking compatibility (shamelessly taken from stackoverflow).

Refractor ForEachDataInputs and NumpyConcatenateOnCustomAxis to ForEach and NumpyConcatenateOnAxis. This fixes #284.

N.B. This features has been rolled back because it created too much problem.

ZipMinibatchJoiner and ZipFeatures

Introduced two classes making use of the ZipDataContainer data structure to join DataContainer instances at different point in the pipeline. ZipMinibatchJoiner, as its name suggest, zips every n-th element of returned minibatches DataContainer together. ZipFeatures is a steps which expects its input to be a DataContainer containing a list of DataContainer and not expected output. ZipFeatures is designed to be used as joinder for FeatureUnion instance where concatenation is not possible or desireable (but could be used as a regular step also).

Other changes

• Introduced Trial function get_best_validation_score() and Trials get_best_trial()
• Fixed a bug in MetaStep and TruncableStep mutate
• Also includes relevant changes in PR #369, so it's safe to close it once this is merged.
• Introduced AssertExpectedOutputIsNotNoneMixin
• Added on_complete call to HyperparamsRepository subscriber at the end of an AutoMl loop
• Implemented SaveBestModelCallback so that each trial may keep its best model stored.
• Added constructor to ToNumpy step (fixes #285)
• Closes #262
• Added joiner argument to ColumnTransformer
• Fixes #469
• De-numpyfied ColumnSelector2d, thus making it compatible with non-contiguous array
• Muting two warning which were way too frequent (step renaming warning and minibatch setting joiner batch size automatically)

[guillaume-chevalier]

@vincent-antaki Thanks, I will review this when the other branch is merged. To speed up the process, you may open a PR in your own fork from this branch to your master and then DM me to review it.

[vincent-antaki]

@guillaume-chevalier This is ready to merge. I've applied and/or answered all your comments.

[pull-checklist[bot]]

New contributor? Ensure you do this

• [ ] Your local Git username is set to your GitHub username, and your local Git email is set to your GitHub email. This is important to avoid breaking the cla-bot and for your contributions to be linked to your profile. If at least 1 contribution is not commited properly, the cla-bot will break.
• [ ] Use the PyCharm IDE with PyTest to test your code. Reformatting your code at every file save is a good idea, using PyCharm's Ctrl+Alt+L shortcut.
• [ ] Sign the Contributor License Agreement (CLA) to allow publishing your contributions under the Apache 2.0 license, in order for everyone to be able to use your code.