axlesubash/wisdm-watch-stat-xgb

WISDM watch statistical XGBoost

CPU classifier for one 5 s, 20 Hz smartwatch IMU window (T=100, C=6) into 18 WISDM activities. The ONNX file is the XGBoost tree head only. Statistical features (104 dims) still run in Python in notsubash/Activity-Recognition.

Not a medical or safety device. Watch windows only. Do not send phone IMU to this bundle.

Metrics to cite

Protocol B, 5-fold GroupKFold on subject_id, 51 subjects, repaired 20 Hz WISDM (UCI 507). Source: docs/reports/protocol_b_watch_stat_xgb.json.

Metric	Value
macro-F1	0.7031
accuracy	0.7013
mean fold macro-F1	0.7027 (std 0.0506)
locomotion group F1	0.9292
posture group F1	0.6606
hand group F1	0.8788
eating group F1	0.8450

Primary metric is macro-F1. The ONNX in this repo is a refit on all windows (one subject held out only for XGBoost early stopping). It is not a GroupKFold fold. Cite the table above, not an export-fit score.

Weak watch class: eating sandwich (L), per-class F1 0.2816. Stairs 0.7028, kicking 0.7831.

Files

File	Role
watch_stat_xgb.onnx	XGBoost classifier (onnxmltools / ONNX 1.22, input [None, 104])
watch_stat_xgb.json	Sidecar (har.onnx.v1): window contract, device, classes
LICENSE	MIT

Config used to train: configs/protocol_b_watch_stat_xgb.yaml (200 trees, max_depth 6, statistical features, 5.0 s window, 1.0 s hop).

Use

Install the package from GitHub, then download this repo so the .onnx and .json sit together:

from pathlib import Path

import numpy as np
from huggingface_hub import snapshot_download

from har.models.export import load_bundle, predict_window

local = snapshot_download("axlesubash/wisdm-watch-stat-xgb")
bundle = load_bundle(Path(local) / "watch_stat_xgb.onnx")

# 100 rows, channels ax, ay, az, gx, gy, gz
window = np.zeros((100, 6), dtype=np.float32)
print(predict_window(bundle, window))

Input contract:

• device: watch
• hz: 20
• shape: (100, 6)
• channels: ax, ay, az, gx, gy, gz

Phone windows are out of contract. This is not transformers.AutoModel.

Training data

WISDM Smartphone and Smartwatch Activity and Biometrics Dataset (UCI 507): 51 subjects (1600-1650), 18 activities (A-S skipping N), watch on the dominant hand, accelerometer and gyroscope. Sessions are repaired onto a shared 20 Hz grid. There are no demographics, so there is no fairness slice. Raw WISDM is not redistributed here.

Limitations

• Subject-independent watch HAR only. Phone statistical XGBoost is 0.3272 macro-F1 under the same protocol and is a different bundle.
• Sandwich, stairs, and kicking are the weaker watch classes.
• Abstain threshold is 0.0 (never abstain) and is uncalibrated.
• Statistical features stay in Python. onnxruntime runs the tree head only.
• No 12-channel phone+watch fusion.

Citation

WISDM (Weiss et al., UCI 507). Training code and protocol: notsubash/Activity-Recognition. Weights in this Hub repo are MIT. The dataset has its own UCI terms.