Learned personal model — local MLP / gradient boost trainer (Phase A+) #3

Open
opened 2026-04-26 21:43:01 -07:00 by pyr0ball · 0 comments
Owner

Local classifier that learns user-specific gestures on top of the rule-based defaults.

Training: scikit-learn MLP or GradientBoostingClassifier. Input: normalized landmark vectors from cf_input.gestures (63-dim hand, or face landmark subset for blink/gaze/head). Train from certified samples in ~/.merlin/training/. Trains in < 5 seconds on CPU for typical datasets (100–500 samples).

Model persistence: joblib dump to ~/.merlin/models/<profile_name>.joblib. MerlinConfig stores active model path.

Inference: replaces / augments rule-based detector in camera loop. Falls back to rule-based if no model trained yet.

BSL boundary: model training and inference code is BSL. Training data (landmark vectors) and model weights are user-owned, stored locally.

Tests: train on synthetic landmark clusters, verify correct class prediction. Verify fallback to rule-based when model missing.

Local classifier that learns user-specific gestures on top of the rule-based defaults. **Training:** scikit-learn MLP or GradientBoostingClassifier. Input: normalized landmark vectors from cf_input.gestures (63-dim hand, or face landmark subset for blink/gaze/head). Train from certified samples in `~/.merlin/training/`. Trains in < 5 seconds on CPU for typical datasets (100–500 samples). **Model persistence:** joblib dump to `~/.merlin/models/<profile_name>.joblib`. MerlinConfig stores active model path. **Inference:** replaces / augments rule-based detector in camera loop. Falls back to rule-based if no model trained yet. **BSL boundary:** model training and inference code is BSL. Training data (landmark vectors) and model weights are user-owned, stored locally. Tests: train on synthetic landmark clusters, verify correct class prediction. Verify fallback to rule-based when model missing.
pyr0ball added this to the Phase A — Camera-Only milestone 2026-04-26 21:43:01 -07:00
Sign in to join this conversation.
No labels
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: Circuit-Forge/raven#3
No description provided.