Label balancing and symmetric behaviors for training classifiers

[SkepticRaven]

Adds:

• label downsampling routine for training classifiers with balanced labels
• GUI checkbox for optionally conducting the balanced training
• balanced training value stored in both trained classifier + project
• random subsetting is hit by random_seed
• bugfix on storing selected gui option for social features

[SkepticRaven]

Current TODO:

• Update all calls of classifier.train to now include new uses_balance parameter

Right now, this option only really works within the gui. classify.py will fail.

[SkepticRaven]

Adding in another new feature: symmetric behaviors (where left-right features can be swapped)

• Checkbox for enabling this feature in the classifier
• Symmetric behaviors will essentially double the training dataset size and force the classifier to rely on asymmetric features less (due to L-R swapping for the doubling)

In addition to that, a massive rework was done to enable this feature to be added

• Lots of new class functions to get feature names without an instantiated class object
• Storing a new value into saved training data
  • min_pose_version: minimum required pose version for the classifier (currently not used in checks, but might be useful)
• New all-k-fold button for running the maximum number of k-fold cross validations in the project
  • Helps with reporting better averages between multiple trainings when a video may be an outlier in performance (but useful for training)
• Added std reporting for k-fold cross validation

[dahhei]

Are all pose versions currently backwards-compatible?

[SkepticRaven]

Features are backwards compatible, but projects/classifiers can only operate on the lowest common feature set. This is the current way projects calculate this number.

Currently no logic is present for decreasing the number passed here... but it could be added (rather than just relying on param1's value).

[dahhei]

I believe that I can't test these changes on my local install of JABS unless the update-reqs is merged first. There is a version from commit-head 49c9b10 on the HPC that I will use to test these updates. 😃

[dahhei]

Cayson initially raised this issue so I tested out the all k-fold cross-validation and it appears as though the top 10 feature labels are incorrect. I trained a Drinking classifier which typically has the distance to Lixit features present. This one does not. The all k-fold does seem to work otherwise though.

[dahhei]

It would be nice if the cross validation iteration value included a "cross validation iteration 14/n", with n being the total number that it will test. 🤠

[SkepticRaven]

Lower pose version projects aren't getting their feature set correctly propagated to the new functions. On a pose_v2 project:

Traceback (most recent call last):
  File "/behavior-classifier/src/ui/training_thread.py", line 56, in run
    features, group_mapping = self._project.get_labeled_features(
  File "/behavior-classifier/src/project/project.py", line 663, in get_labeled_features
    column_names = features.get_feature_column_names(
  File "/behavior-classifier/src/feature_extraction/features.py", line 539, in get_feature_column_names
    column_names = self.get_feature_name_vector(pose_version=self._pose_version, use_social=use_social, extended_features=self._extended_features)
  File "/behavior-classifier/src/feature_extraction/features.py", line 494, in get_feature_name_vector
    assert pose_version >= 5
AssertionError

[SkepticRaven]

The commits above are a rebase merge with main (based on the other recent merge into main: #27)

[SkepticRaven]

The feature rewrite to propagate names alongside values fixed the pose_v2 bug.
This branch should be ready to merge (after updating the appropriate version numbers)!