Bug for inference

[v3551G]

I don't know if you have tried it, but I have encountered a lot of problems in the process of converting it to ONNX format and reasoning, because the code contains a lot of custom operations, and so on

[ChristophReich1996]

Hi @v3551G, thanks for your interest in our code! Yes, converting to ONNX might be problematic, I never tried the conversion. While we utilize a custom CUDA implementation of the PAU, we also have a pure PyTorch implementation added as a fallback option (see). Maybe try to only import the PAU PyTorch implementation here. This might enable converting the model to ONNX. Let me know if this helps!

[hhuiwang]

+1