search

NVlabs / PWC-Net

PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume, CVPR 2018 (Oral)
Other
1.63k stars 357 forks source link

Inference time on pytorch version #129

Open MiaoJieF opened 2 years ago

MiaoJieF commented 2 years ago

I use the pytorch version of pwcnet to test the inference time, the image resolution is 1024 x 512, the graphics card used is RTX3060, and the inference time is 0.08s. There is a big gap with the reasoning time in the paper. Is this because the pytorch version is slower than the caffe version?

andersonhusky commented 1 year ago

I use the pytorch version of pwcnet to test the inference time, the image resolution is 1024 x 512, the graphics card used is RTX3060, and the inference time is 0.08s. There is a big gap with the reasoning time in the paper. Is this because the pytorch version is slower than the caffe version?

hi, i'm having trouble when I configure the python enviroment of the torch version, could you please teach me how to configure the enviroment of the torch version?

MiaoJieF commented 1 year ago

You can refer to the RAFT environment requirements for configuration.

andersonhusky commented 1 year ago

You can refer to the RAFT environment requirements for configuration

I'm sorry, what is "the RAFT environment "? I follw the steps in Torch/readme.md, but it seems like cuda 11.2 of 3060 doesnt match with the cuda the code need..

MiaoJieF commented 1 year ago

1.RAFT is a well-known optical flow estimation algorithm which is also open source. 2.The cuda version I use is 11.1. The Pytorch version is 1.9.1. 3.Also, see your question under Flowformer. That's due to a lack of memory on the GPU, try using the lightweight version provided by the author.

andersonhusky commented 1 year ago

1.RAFT is a well-known optical flow estimation algorithm which is also open source. 2.The cuda version I use is 11.1. The Pytorch version is 1.9.1. 3.Also, see your question under Flowformer. That's due to a lack of memory on the video card, try using the lightweight version provided by the author.

Thanks! You saved my day!

andersonhusky commented 1 year ago

1.RAFT is a well-known optical flow estimation algorithm which is also open source. 2.The cuda version I use is 11.1. The Pytorch version is 1.9.1. 3.Also, see your question under Flowformer. That's due to a lack of memory on the GPU, try using the lightweight version provided by the author.

Hi, sorry to bother you again. I have successfully configure the enviroment of RAFT, the problem is RAFT is based on python3, and I still cant run pwcnet-torch in this enviroment successfully... by the way, here is my email: hogfeg@zju.edu.cn, if you feel inconvenient to explain here, we can communicate via email, thanks for you generous help again！