Optimize Classification Performance (~2x Faster)

Summary

This PR optimizes classification performance by eliminating redundant classifier calls during prediction. Both GUI and command-line classification are now approximately 2x faster with identical results.

Performance Improvement

Problem

During classification, the code was calling both predict() and predict_proba() separately on the same data:

predictions = classifier.predict(data, frame_indexes)
probabilities = classifier.predict_proba(data, frame_indexes)

This caused the classifier to run twice per identity because predict() internally computes probabilities and then thresholds them at 0.5. For large pose files with multiple identities, this resulted in unnecessarily slow classification.

Solution

Call predict_proba() once and derive predictions using np.argmax():

# Get probabilities for all classes
probabilities = classifier.predict_proba(data, frame_indexes)

# Derive predictions by taking argmax (class with highest probability)
# This is equivalent to predict() but avoids duplicate computation
predictions = np.argmax(probabilities, axis=1).astype(np.int8)

The np.argmax(probabilities, axis=1) returns the index of the class with the highest probability for each frame, which for binary classification is 0 (not behavior) or 1 (behavior) - exactly what predict() would return.

Files Changed

• src/jabs/ui/classification_thread.py - GUI classification in the main JABS application
• src/jabs/scripts/classify.py - Command-line jabs-classify tool

Correctness

✅ Results are identical - np.argmax(predict_proba()) produces the exact same predictions as predict()

✅ No breaking changes - All existing code continues to work

✅ Tested - Classification produces identical results with the optimization