Onnx slower than pytorch

[wingvortex]

Hi,

I tested an onnx model converted (by export_onnx.py) from a torch model (trained with config rtdetr_r50vd_6x). The onnx model is always much slower than the torch model, no matter the CPU, GPU, or batch process. Why does this happen?

[lyuwenyu]

It may relate to the ONNX inference engine, that is not as good as torch (e.g some op).