A conversion pipeline turns a model into a version suited to a target runtime or accelerator. For small-box deployments, build and validate a separate pipeline for each vendor: calibration can produce different activation ranges and shift accuracy even when vendors use the same model and representative data. section 11.4
Account for these deployment costs:
- Maintain vendor-specific conversion scripts and runtime binaries. section 11.4
- Repeat INT8 calibration for each vendor compilation; calibration ranges can differ and accuracy can shift by 1–3 percentage points. section 11.4
- Check accelerator operator support; unsupported models may need an SBC CPU-only path. section 11.4
One model, four targets
Deploying the same EfficientNet to Jetson, OpenVINO, Coral, and Hailo requires four conversion scripts and four runtime binaries, with calibration discipline repeated for each vendor. section 11.4
