SIAM RPN on hailo8l

General
1

Has anyone already worked with quantizing SiamRPN for Hailo-8? Are there perhaps any guides available?

2

Hey @An_ti11

we don’t officially support SiamRPN, but here’s a basic guide to get started with any custom model on Hailo-8:

Setup and Compile SiamRPN

  1. Export your PyTorch model to ONNX
    Use torch.onnx.export() with proper input/output names.

  2. Translate the ONNX model using Hailo SDK
    Use ClientRunner().translate_onnx_model(...) with the correct input shape.

  3. Apply custom quantization parameters

    • Use quantization_param(...) for precision per layer.
    • Set quantization_groups and force_range_out as needed.

  4. Run model calibration
    Use representative data and call model_optimization_config(...).

  5. Run model compilation
    Use ClientRunner().compile()

Tips for SiamRPN

  • Ensure cross-correlation (group conv) layers export correctly to ONNX.
  • Watch for reshape/transpose issues – they often break shape inference.
  • Quantization is tricky in dense and upsampling layers – tune carefully.