Deploy to Edge Device
Using the Model on the Board
Once deployment is complete, use the ShimetaPi Software Repository tools to run inference:
# Image inference
shimeta_infer /mnt/your_model.zip image.jpg
# Real-time camera inference
shimeta_camera /mnt/your_model.zipModel Format Reference
| Export Format | File Extension | Usage |
|---|---|---|
| XMM (GK7206) | .xmm.zip | Contains neuron_network.xmm and model_meta.txt |
| RKNN (RK3588) | .rknn | Loaded directly via RKNN Toolkit |
FAQ
Q: How do I deploy the exported model to the board? A: Download the model file in the appropriate format (RKNN/XMM), transfer it to the board via SCP, then load it following the board SDK documentation.
Q: What's the difference between quantized and non-quantized export? A: Quantization (INT8) reduces model size and speeds up inference, but accuracy may be slightly lower.
Q: How many classes can one model detect? A: YOLO has no hard limit on class count in theory, but more classes make training harder. Recommended: no more than 20 classes per model.