Transfusion复现

CVPR 2022
paper: https://arxiv.org/abs/2203.11496
code: https://github.com/XuyangBai/TransFusion/tree/master

这是我在配环境中遇到相当难复现的一篇文章，特此记录下能够正确运行的版本

• Ubuntu20.04
• CUDA=11.3
• PyTorch: 1.10.1
• TorchVision: 0.11.2
• MMCV: 1.4.0
• pip install mmcv-full==1.4.0 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10/index.html

或者下载好离线安装
https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/mmcv_full-1.4.0-cp38-cp38-manylinux1_x86_64.whl
pip install mmcv_full-1.4.0-cp38-cp38-manylinux1_x86_64.whl

• mmdet==2.11.0

mmdet=2.20.0报错，没有build

• numpy=1.19.5

解决这个报错的issue: ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 88 from C header, got 80 from PyObject
https://stackoverflow.com/questions/66060487/valueerror-numpy-ndarray-size-changed-may-indicate-binary-incompatibility-exp
修改pycocotools的版本

• pycocotools=2.0.1

• pip install -v -e .
  编译报错，按照这个issue修改 mmdet3d/ops/voxel/src/scatter_points_cuda.cu中的一段代码

  // coors_map.index_put_(coors_id_argsort, coors_map_sorted);
  coors_map.index_put_({coors_id_argsort}, coors_map_sorted); //! dxg

然后要从lidar-only的模型开始训

python -u tools/train.py configs/transfusion_nusc_voxel_L.py