PyTorch and ONNX to RKNN in three steps
A conversion tool for Rockchip NPUs with a web UI and no CLI. Supports the full YOLOv8 family (Det / Seg / Pose / OBB), ResNet, and RetinaFace, with INT8 quantization, a calibration set builder, and real-time SSE logs.
What it does
Each capability maps to a concrete everyday problem.
Three-step card UI
Pick a model, set options, start converting — each step documents itself with sensible defaults.
Full YOLOv8 family
Supports all four YOLOv8 tasks — Det, Seg, Pose, OBB — plus ResNet and RetinaFace presets.
Two conversion paths
Convert from PyTorch via TorchScript, or straight from an existing ONNX — whichever you have.
INT8 quantization
Built-in INT8 quantization gives higher NPU throughput than FP16 with controlled accuracy loss.
Calibration set builder
Automatically samples a representative calibration set from your data, or point it at a folder.
Real-time SSE logs
Each conversion step streams over SSE so you can see exactly where it stalls.
How it works in practice
Not a feature list — a walk through getting something done.
Choose a model, configure quantization, convert
First, upload or point to the model file and the tool detects the task type with recommended settings.
Then confirm input size, quantization mode, and calibration source, and start converting with live logs.
When it fails, the log shows which layer broke
SSE pushes each step's result, preserving full context on failure so you can compare against known workarounds.
Preview the resulting RKNN graph with Netron right in the page to verify the conversion immediately.
Design trade-offs
Deterministic work stays in code; judgement stays with you.
Runs locally
The web UI runs on your machine; model files never go to a third-party server.
Graph preview
Netron integration lets you confirm the graph matches expectations after conversion.
Visible progress
Long steps report their stage and progress, so you are never left guessing.
Ready-made presets
Common models come as presets, cutting down trial and error.
What you might want to know first
What environment does it need?
A local Python environment with RKNN Toolkit2. The tool provides the web UI that orchestrates conversion and shows logs.
Which models are supported?
The full YOLOv8 family, ResNet, and RetinaFace, plus other models via the ONNX path.
How do I get it?
See the GitHub repository for installation and usage.
Start using it now
Free to download, runs locally. Your data stays on your machine.
This is a desktop app — open this page on your computer to download.
Other formats
SHA256 checksum
Provided as-is. The app does not collect usage data.