qriton/hlm-micro-gesture-v2-uwave

HLM-Micro Gesture v2 - UWave

HLM-Micro Gesture v2 is a small polynomial-Hopfield time-series classifier trained on the UWaveGestureLibrary real accelerometer gesture corpus.

Results

Field	Value
Parameters	183,933
Classes	8 gestures
Reported test accuracy	94.19%
Best checkpoint	epoch 18 of 30
Input	3-axis accelerometer, resampled to a 32-channel feature layout

Reference points from common UWave time-series classifiers are in the high-80s to mid-90s depending on method and split. This checkpoint is a compact research model, not a production wearable system.

Files

File	Purpose
model.pt	Sanitized model-only PyTorch checkpoint
config.json	Public architecture, task, classes, and metric metadata
metrics.jsonl	Training/evaluation metrics from the local run

Intended Use

• Research on compact polynomial-Hopfield classifiers.
• Edge/time-series gesture-recognition experiments.
• Baseline for replayable per-inference audit work.

Limitations

• Evaluated on UWaveGestureLibrary, not on arbitrary users or devices.
• No safety, medical, ergonomic, or production-use certification.
• Real deployments need sensor calibration, broader classes, and live-device validation.