Flip classifier app uses cleaned frames as training data, whereas extraction performs inference on unfiltered data.

[jackransomlovell]

moseq2_app.flip.flip_classifier_tool.train_and_evaluate_model() trains on cleaned frames, whereas moseq2_extract.extract.proc.get_flips() performs inference on uncleaned frames.

Expected Behavior

RF Flip classifiers should perform inference on the same type of frames (cleaned or uncleaned) they are trained on.

Current Behavior

RF Flip classifiers perform inference on uncleaned frames, while being trained on cleaned frames.

Possible Solution

Depends on how the default flip classifiers were trained.

Steps to Reproduce (for bugs)

1. self.corrected_dataset is cleaned frames and is later used as self.x in training function
2. Inference is performed on non-cleaned frames

Screenshots

Context

Your Environment

• Version/GitHub branch used:
• Device: [e.g. iPhone6]
• OS: [e.g. iOS8.1]
• Link to your project:

[jackransomlovell]

The kernel size for the tail filter is also hardcoded, and should instead be defined by the user (and ideally passed from the config file).

fixed with c84e1cd

[jackransomlovell]

fixed with dattalab/moseq2-extract@ddb41c6

[jackransomlovell]

allow users to define all cleaning parameters used, ideally these would just be called from the config file in the notebook example
94545e7

[jackransomlovell]

cleaning params implemented in notebook 78a698d