Custom Models (DXNN SDK)
Take a model the Sixfab Model Zoo does not ship: export it to ONNX, compile it to a .dxnn file with DX-COM, and deploy it on the DEEPX NPU. Covers the two-machine workflow, the full config.json reference, and the Ultralytics export path.
The Sixfab Model Zoo covers the common vision models. This section is for everything else: your own trained weights, an architecture the zoo does not ship, or an Ultralytics model you want to accelerate on the DEEPX NPU. Compilation happens once on an x86_64 Ubuntu host; the resulting .dxnn file runs unchanged on Sixfab AI HAT+, the Edge AI Expansion Board, and ALPON X5 AI.
- Deployment Workflow — the end-to-end path: the two-machine model, compiler host requirements, ONNX export rules, INT8 calibration, and deploying the compiled artifact to the target device.
- DX-COM Configuration Reference — every
config.jsonkey: input shapes, calibration settings, PPU post-processing for YOLO, DXQ accuracy recovery, and the supported preprocessing operations, with copy-ready examples. - Sixfab × Ultralytics Acceleration Path — the three ways to get an Ultralytics YOLO model to DEEPX format: native export, the Ultralytics Platform, and local DX-COM.
Before you compile, check that the architecture and its operators are supported: Supported Models.
Updated 6 days ago
Did this page help you?