- Create conda environment with following command:
cd Cylinder3D-Spconv-2.2-PyTorch-Lightning
conda env create --name envname --file=environment.yml
- Install PyTorch
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
- Install torch-scatter
pip install --no-index torch-scatter -f https://data.pyg.org/whl/torch-1.12.1+cu113.html
- Install spconv
pip install spconv-cu114==2.2.6
- Install other dependencies
pip install -r requirements.txt
- modify the
config/semantickitti_pl.yaml, including data path, log path (tensorboard log and model checkpoint), batch size, ddp training GPU device number - train the network by running
CUDA_VISIBLE_DEVICES=0,1 python train_cylinder_asym_pl.py
- evaluate the trained network by running
CUDA_VISIBLE_DEVICES=0,1 python train_cylinder_asym_pl_eval.py(can eval on single or multiple GPUs)
- 40 epochs
- 0.000707 base LR with sqrt_k scaling rule (equals to original 0.001 at batchsize = 2, equals 0.00489 at batchsize = 24)
- AdamW with Weight Decay 0.001
- CosineDecay Schedule
| GPU Type | Single GPU | 4 GPUs |
|---|---|---|
| A100 | 60 mins/epoch | 25 mins/epoch |
If you find our work useful in your research, please consider citing the original paper paper and the repo:
@article{zhu2020cylindrical,
title={Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation},
author={Zhu, Xinge and Zhou, Hui and Wang, Tai and Hong, Fangzhou and Ma, Yuexin and Li, Wei and Li, Hongsheng and Lin, Dahua},
journal={arXiv preprint arXiv:2011.10033},
year={2020}
}
@software{githubGitHubJaywu109Cylinder3DSpconv22PyTorchLightning,
author = {Wu, Dai-Jie},
title = {{G}it{H}ub - jaywu109/{C}ylinder3{D}-{S}pconv-2.2-{P}y{T}orch-{L}ightning --- github.com},
url = {https://github.com/jaywu109/Cylinder3D-Spconv-2.2-PyTorch-Lightning},
year = {2023}
}
We thanks for the opensource codebases, Cylinder3D and Cylinder3D-updated-CUDA