aselimc Edge Deployment
Runtime Selection
| Hardware | Runtime | Format |
|---|---|---|
| Jetson (NVIDIA) | TensorRT | .engine |
| Mobile (ARM) | TFLite | .tflite |
| Intel CPU/GPU | OpenVINO | .xml/.bin |
| Apple Silicon | CoreML | .mlmodel |
| Any CPU | ONNX Runtime | .onnx |
Containerized Edge
FROM nvcr.io/nvidia/l4t-tensorrt:r8.5.2-runtime
COPY model.engine /models/
COPY app.py /app/
CMD ["python", "/app/app.py"]
Use K3s for orchestration, Balena for fleet management.
OTA Model Updates
- Version models with hash + timestamp
- Download new model in background
- Validate checksum, test inference on canary input
- Atomic swap (symlink switch)
- Rollback on failure
Telemetry
MQTT for lightweight device-to-cloud: detections, health, metrics.
Key Libraries
TensorRT, TFLite, ONNX Runtime, MQTT (mosquitto), AWS IoT Greengrass