search

hmorimitsu / ptlflow

PyTorch Lightning Optical Flow models, scripts, and pretrained weights.
Apache License 2.0
250 stars 33 forks source link

TensorRT of rapidflow #65

Closed ylab604 closed 3 months ago

ylab604 commented 3 months ago

Hello! Thank you for great work about rapidflow ICRA2024 I remember what you explained in the poster. And write this email while testing rapidflow. I am currently changing rapidflow to onnx and changing it to TensorRT and experimenting with orin nx. But it doesn't seem to be robustly changed to TensorRT. The output value is only a numpy array of constant values. Do you know anything about this? How did you experiment with converting to TensorRT?

Once again, thank you so much for revealing such good research.

ylab604 commented 3 months ago

I changed rapidflow_it12 to onnx as per the readme and changed it back to TensorRT. It has been confirmed that onnx inference works, but the result of TensorRT only comes out as a constant.

rapidflow_it12.ckpt ==> onnx (O) onnx ==> TensorRT (X)

i have a Question is your Tensorrt ver : 8.6?

hmorimitsu commented 3 months ago

Hello, thank you for your report.

I pushed a new update with some modifications to the rapidflow code and a simple script to compile to TensorRT.

Please check rapidflow's readme for some instructions. I tested it on my computer, not on a Jetson. It seems to work on the sample images, but I did not test it very extensively. If there is still some problem, please let me know.

The versions in my test were: torch 2.2.2+cu121 torch_tensorrt 2.2.0 tensorrt 8.6.1

Best regards

ylab604 commented 3 months ago

thank you for kindness i will check and write the result soon

ylab604 commented 3 months ago

it works at orin nx jetpack 6.0 onnx opset 17 tesnsorrt 8.6.2 thank you alot!! best regards

I'll say thank you when I see you at the next conference. Let's have a meal that time.

hmorimitsu commented 3 months ago

Glad to know it is working! Yes, looking forward for our next meeting.