search

sair-lab / AirLine

AirLine: Efficient Learnable Line Detection with Local Edge Voting (IROS 2023)
https://sairlab.org/airline
BSD 3-Clause "New" or "Revised" License
66 stars 3 forks source link

Questions about the inference code （need model.eval()） #8

Open BaofengZan opened 6 months ago

BaofengZan commented 6 months ago

I use the ‘Airline_demo.py’ to test the images and the original file can get normal results. But when I add Premodel.eval(), the result is 0（other parameters remain the same）.

without model.eval() image

with model.eval() image

Lx017 commented 6 months ago

hmmmm, I had not tested this though, this may affect some normalization but I am not sure. if no .eval() works then I recommend to just leave it so?

BaofengZan commented 6 months ago

When I export onnx using torch.onnx.export, the bn layer exports in eval mode by default， If the onnxruntime inference is used, the result is incorrect. But when I export the bn layer in TRAINING mode， when converting to tensorrt, it will prompts that the BN layer is in TRAINING mode, which makes it impossible to convert. So I think I still need to check this issue.

Error when converting tensorrt。Official Answers： https://github.com/NVIDIA/TensorRT/issues/3457#issuecomment-1817477998 image

Lx017 commented 6 months ago

yeah sorry, that is indeed something we missed...

BaofengZan commented 6 months ago

Thank you, I look forward to you being able to solve this problem.