Partial training

[albertcthomas]

On behalf of @kegl

This PR adds the possibility of partial training to the workflow by loading a previous pickled model

• [x] add command-line option --partial-train
• [x] add optional prev_model to train_workflow in classifier, regressor, feature_extractor_classifier, and feature_extractor_regressor
• [x] titanic_old test with partial training
• [x] air_passenger_old test with partial training
• [x] decouple pickling and unpickling
• [x] test pickle_model and unpickle_model
• [x] name change from (un)pickle_model to (un)pickle_trained_workflow to be more precise

[kegl]

Ping @agramfort and @rth , if someone would like to review from the Saclay side. It went through internal review on ours, @albertcthomas will comment tomorrow. The two reasons we decided to do this:

1. We want to do warm start in hyperopt.
2. We are partially training the system models in our rl_simulator application (generative_models branch). But it can also be used from the command line if someone wants to train, the stop and check the errors, then start from the partially trained model. Especially in the development phase it could be handy.

[albertcthomas]

Travis is missing because of the migration from travis-ci.org to travis-ci.com. I don't think I have the rights to add the repo to travis-ci.com.

[rth]

Should we just migrate to GH Actions? Those are enabled by default, and the CI setup is fairly straightforward in this repo.

[albertcthomas]

Should we just migrate to GH Actions? Those are enabled by default, and the CI setup is fairly straightforward in this repo.

This might be better indeed.

[kegl]

Should we just migrate to GH Actions? Those are enabled by default, and the CI setup is fairly straightforward in this repo. Would be great!

[rth]

Could you please merge master in to enable CI?

[codecov[bot]]

Codecov Report

Merging #287 (3cfd1fc) into master (aee6274) will increase coverage by 0.15%. The diff coverage is 94.56%.

@@            Coverage Diff             @@
##           master     #287      +/-   ##
==========================================
+ Coverage   82.97%   83.12%   +0.15%     
==========================================
  Files         137      137              
  Lines        4680     4736      +56     
==========================================
+ Hits         3883     3937      +54     
- Misses        797      799       +2     

Impacted Files	Coverage Δ
..._old/submissions/starting_kit/feature_extractor.py	100.00% <ø> (ø)
rampwf/utils/testing.py	100.00% <ø> (ø)
rampwf/utils/submission.py	87.56% <86.48%> (+0.29%)	:arrow_up:
...ssengers_old/submissions/starting_kit/regressor.py	100.00% <100.00%> (ø)
...titanic_old/submissions/starting_kit/classifier.py	100.00% <100.00%> (ø)
rampwf/tests/test_kits.py	97.10% <100.00%> (+0.08%)	:arrow_up:
rampwf/utils/__init__.py	100.00% <100.00%> (ø)
rampwf/utils/cli/testing.py	63.41% <100.00%> (+0.91%)	:arrow_up:
rampwf/utils/tests/test_submission.py	100.00% <100.00%> (ø)
rampwf/workflows/classifier.py	100.00% <100.00%> (ø)
... and 3 more

---

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 aee6274...3cfd1fc. Read the comment docs.

[kegl]

@agramfort if no objections from the Saclay side, we'll merge this next week.