v0.0.3: merge staging into main

[levinwil]

v0.0.3

Added Features

1. Added n_estimators to LifelongClassificationForests add_{task, transformer} functions. This allows each task to have a different number of trees.
2. Previously, the classes in forest.py used finite_sample_correction with the only possible value of kappa being 1. We added the ability to set any kappa.

Bug Fixes

1. Previously, the default values of all model parameters to LifelongClassification{Forest, Network} add_{task, transformer} functions were None. The value of None indicated to use the default model parameter specified in the instantiation of the LifelongClassification{Forest, Network}. But, this meant that a user would be unable to train a new tree transformer to purity (by setting max_depth = None) unless default_max_depth was None. This is an undesirable restriction. So, we changed the default value of all model parameters to LifelongClassification{Forest, Network} add_{task, transformer} functions to the string "default" - this indicates to use the default model parameter specified in the instantiation of the LifelongClassification{Forest, Network}.
2. Updated selection of {transformer, voter, decider} data to be without replacement.
3. Previously, we were encountering issues when the voters across bags would output posteriors of different lengths (caused by only training on a subset of the classes). We manually fixed these to ensure that all voters across bags output the same number of classes, by appending 0's to missing classes in the voter posterior estimates.
4. Type checked all inputs X, y to {network, forest}.py classes.

Documentation Fixes

1. Previously, many of the docs did not render correctly in Sphinx because they were formatted incorrectly. We fixed this formatting error and now all documentation renders correctly in Sphinx (and thus the web docs).

Presentation Changes

1. transformer_voter_decider_split was changed to network_construction_proportion in network.py
2. Minimized documentation and Python Package to only expose well-documented, tested functions that are described in the paper

[PSSF23]

TODO: fix UncertaintyForest and optimize figure 2 tutorial #365

[levinwil]

@PSSF23 What do you think about merging this into main now that UF has been fixed (assuming tests pass)?

[PSSF23]

@levinwil Yeah, it should be ready now as progress towards #1.

[levinwil]

@PSSF23 I agree it should be merged.

But I don't understand your second comment. How is this progress towards #1 ?

[PSSF23]

@levinwil the package is becoming more sklearn compliant I believe? So each version would be progress towards #1?

Or is the issue complete already? We do have proglearn on pip.

[levinwil]

@PSSF23 I think there are many goals. One is becoming more sklearn compliant. I could be wrong but I don't think anything in this PR makes it more sklearn compliant - but that is beside the point. I understand the comment now.

As for your question, it's not sklearn compliant yet. All of the classes that inherit from an sklearn base class need to be be compliant as an sklearn estimator - it needs to pass the check_estimator function from sklearn. A more comprehensive tutorial can be found at https://scikit-learn.org/stable/developers/develop.html. The TLDR is that it's not there yet - we haven't completed #1 .

[jdey4]

@levinwil Unfortunately I don't have write access. @jdey4 please accept #346 and #362 before merging this one, thank you!

Is it solved? Do you guys need me for anything?

[PSSF23]

@jdey4 the above PRs are merged already. Please check this one for any potential issue. Thanks!