im using an oriented object detetion yolo model, when i run runner.translate_onnx_model i afraid as it mention hailo translation not support trig operation. if i chose after the yolo head later it might lose some of the valuable postprocess operation any suggestion
The problem you’re facing stems from the Hailo ONNX model translation process, which doesn’t support certain operations - specifically trigonometric functions like Cos and Sin, as well as a broadcasting shape issue. Hailo’s dataflow compiler and runtime software (HailoRT) may not directly support these operations, so you’ll need to find alternative solutions.
Here are some approaches to address this:
- Preprocess Trigonometric Operations: Move Cos and Sin calculations outside the model. Perform these operations during data preprocessing instead.
- Create a Custom Post-Processing Step: Implement unsupported operations (like trigonometric functions and shape manipulations) as a post-processing step on the host CPU. This allows you to run the main inference on Hailo hardware while handling unsupported operations separately.
- Adjust Model Architecture: Replace unsupported operations with approximations or alternatives that the Hailo Dataflow Compiler can handle.
- Use Recommended End Nodes: Try parsing the model again using the suggested end node: “/model.22/df1/Reshape_1”. This may help avoid some unsupported nodes.
Regards