Closed fangli333 closed 1 year ago
I haven't tried reproducing the metrics with cuda 11.7. The conda environment I used is largely based on BANMo's conda environment, which uses cuda 11.3.
That being said, I think the evaluation code would run just fine with a conda environment that uses cuda 11.7. To create such a conda environment, you could erase the following 6 lines in misc/totalrecon-cu113.yml
, create the conda environment using this modified .yml file, and then manually install their cuda117 equivalents:
https://github.com/andrewsonga/Total-Recon/blob/5cdfbb98f64a8ee80b2512605c8ff38fdcff0528/misc/totalrecon-cu113.yml#L13-L14 https://github.com/andrewsonga/Total-Recon/blob/5cdfbb98f64a8ee80b2512605c8ff38fdcff0528/misc/totalrecon-cu113.yml#L62-L63 https://github.com/andrewsonga/Total-Recon/blob/5cdfbb98f64a8ee80b2512605c8ff38fdcff0528/misc/totalrecon-cu113.yml#L68-L69
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia; conda install cudatoolkit-dev=11.7 -c conda-forge
Please let me know if you run into any conflicts as a result of using cuda117, or if you have any other general questions.
Another thing you probably would need to change is how you install detectron2. Currently, the README instructs you to install detectron2 from a pre-built wheel, but these pre-built packages are only available up to cuda 11.3. Therefore, for cuda 11.7, you have install detectron 2 from source as shown below:
BEFORE python -m pip install detectron2 -f \ https://dl.fbaipublicfiles.com/detectron2/wheels/cu113/torch1.10/index.html
AFTER python -m pip install -e third_party/detectron2
Closing this issue due to inactivity.
Hi, can I reproduce it on cuda117?