Export Model
After training is complete, convert your model to a format runnable on edge hardware.
Supported Export Formats
| Format | Target Platform | Use |
|---|---|---|
| PT | — | PyTorch raw weights — continue training or use for Python inference |
| ONNX | — | Universal intermediate format — can be converted to other platform formats |
| RKNN | RK3588 | Rockchip NPU accelerated inference |
| XMM | GK7206 | Goke NPU accelerated inference |
Export Steps
- After training completes, click "Export Model" on the training detail page
- Select target format (RKNN / XMM)
- Click "Start Export"
- Wait for conversion to finish (typically 2-5 minutes), then click Download
Model Conversion Pipeline
Training output (.pt) → ONNX (.onnx) → RKNN (.rknn) ← for RK3588
→ XMM (.xmm) ← for GK7206Warning
RKNN and XMM exports depend on specific toolchains on the GPU server. If export fails, contact support to verify the environment.
Quantization
| Option | Size | Speed | Accuracy |
|---|---|---|---|
| No quantization (FP32) | Large | Baseline | Full |
| INT8 quantization | ~4× smaller | ~2× faster | Slightly lower |
The calibration dataset for quantization is taken from the training set.
Download
Exported models can be downloaded directly from the platform. Boards supported by the ShimetaPi Software Repository can import and use them directly.