search

NVIDIA / flownet2-pytorch

Pytorch implementation of FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
Other
3.12k stars 739 forks source link

Support for CUDA 6.5 and SM 3.0? #208

Open ulatekh opened 4 years ago

ulatekh commented 4 years ago

Feel free to tell me to buzz off for having such an old video card, but hopefully you won't...I have a Quadro K600 card, which was fine for my needs until now, and thanks to international catastrophe, I am not in a position to buy a new one.

I'm running Fedora Core 30, and I've installed the latest version of CUDA that supports my video card via nVIDIA's dnf package repository.

I'm just trying to build/run https://github.com/zllrunning/video-object-removal , but I'm unable to build the first dependency, i.e. inpainting/lib/channelnorm_package. It dies with this error:

$(TORCH)/lib/include/ATen/Half.h:23:23: cuda_fp16.h: No such file or directory

Apparently my ancient CUDA doesn't have that header file. Conditional compilation based on CUDA version would be helpful, but is probably a lot of work. To make matters more complicated, the video-object-removal project says it only works with Torch 0.4.0.

ulatekh commented 4 years ago

Please delete this issue...at the time, this was the only "torch" related project I could find.