search

HKUST-Aerial-Robotics / Stereo-RCNN

Code for 'Stereo R-CNN based 3D Object Detection for Autonomous Driving' (CVPR 2019)
MIT License
686 stars 177 forks source link

Support CUDA > 9.0... #16

Closed thully closed 5 years ago

thully commented 5 years ago

With CUDA 9.2 installed, I am unable to build the stereo-rcnn project. Following the instructions on the wiki installs a version of PyTorch that is built for CUDA 8.0 (and complains about missing CUDA 8.0 libraries when building). Changing the cuda80 argument to cuda92 makes no difference, as there seems to be no CUDA 9.2 version of PyTorch 0.3.0 (there is for later versions, but the wiki suggests these aren't supported).

The only thing that ultimately worked was downgrading to CUDA 9.0 and using the cuda90 argument when installing PyTorch. While I could run make.sh by using the cuda90 PyTorch with CUDA 9.2, I couldn't actually run the demo without downgrading CUDA itself.

It seems this package needs to be updated to support newer versions of PyTorch and/or CUDA. As it stands, CUDA 9.0 is not officially supported on Ubuntu 18.04 (the current LTS), or by PyTorch 0.3.0. In the meantime, it would help to document the incompatibility with recent CUDA versions...

PeiliangLi commented 5 years ago

OK, I will update to pytorch 1.0 soon.

congyingTech commented 5 years ago

hello, cuda90 is still not support...

PeiliangLi commented 5 years ago

🚩 This weekend!

PeiliangLi commented 5 years ago

Pytorch 1.0 is supported now. Please checkout to https://github.com/HKUST-Aerial-Robotics/Stereo-RCNN/tree/1.0 for the details. Feel free to post any issue here.