Inference & Dashboard¶

Live inference¶

from host.serial_reader import SerialReader
from host.feature_extraction import FeatureExtractor
from host.live_inference import LiveClassifier

classifier = LiveClassifier("models/worker_har_best.pt")
extractor = FeatureExtractor()

with SerialReader("/dev/ttyACM0") as reader:
    for frame in reader:
        feats = extractor.add_frame(frame)
        if feats is not None:
            label, conf = classifier.predict(feats)
            print(f"{label} ({conf:.2%})")

EMA smoothing¶

LiveClassifier applies exponential moving average smoothing to the model logits with decay factor 0.6. This reduces jitter between consecutive windows while keeping the response time under ~1 second.

classifier.reset()           # clear EMA state (e.g., between sessions)
top3 = classifier.top_k(feats, k=3)  # top-3 labels with confidence

Offline inference (recorded file)¶

from host.data_recorder import DataRecorder
from host.feature_extraction import FeatureExtractor
from host.live_inference import LiveClassifier

data = DataRecorder.load("session_001.h5")
classifier = LiveClassifier("models/worker_har_best.pt")
extractor = FeatureExtractor()

# Iterate over frames in the HDF5 file
for i in range(len(data["timestamps"])):
    # reconstruct SensorFrame from HDF5 slices
    feats = extractor.add_frame(frame_from_hdf5(data, i))
    if feats is not None:
        label, conf = classifier.predict(feats)

Dashboard¶

Launch with:

cd wearable_har_v2
.venv/bin/python -m streamlit run host/dashboard.py

Opens at http://localhost:8501.

Panels¶

Panel	Description
Stick figure	3D stick figure from IMU orientations (wrists, elbows, chest, thigh). Rotates live.
Activity gauge	Current classification with confidence bar. Shows top-3 labels.
EMG bars	Forearm flexor + lumbar erector activation, 0–100%
FSR heatmap	2×2 foot pressure grid (heel/toe × left/right)
Timeline	30-second scrolling activity timeline with color-coded labels
FPS counter	Streamlit rerun rate

Features¶

Auto-start: opens SerialReader and starts streaming on page load
Model selector: dropdown to pick a .pt model file
Port selector: dropdown for serial port (auto-detects /dev/ttyACM*)
Dark mode: Streamlit native, toggle in settings menu

Health check¶

curl -s http://localhost:8501/healthz
# → "ok"

Performance notes¶

The dashboard target is 10 FPS (100 ms rerun). Data arrives at 100 Hz; the dashboard decimates by processing every Nth frame. If the UI lags, reduce the number of stick-figure samples or close the timeline panel.