Skip to content

Latest commit

 

History

History
120 lines (94 loc) · 5.47 KB

INSTALL.md

File metadata and controls

120 lines (94 loc) · 5.47 KB
Raw

Installation

We provide three ways to install robot: 1) a partial installation 2) a custom installation and 3) using Docker.

Partial installation

If you only work on registration tasks, we can use following steps to install RobOT.

  • Important. Before the installation, make sure the cuda-toolkit is installed. You can check if it is installed via "nvcc --version" in the terminal (see step 0 if nvcc is not yet installed). The cuda compiler version it shows may be different from your cuda driver version shown at "nvidia-smi". Please make sure that torch, and torch_scatter are installed under the same cuda version as the one of nvcc. (Note that if your nvcc version is 11.2 as pytorch and torch_scatter of version 11.2 are not released, you can install any available version compiled with cuda 11.*)

We assume all the following installation is under a conda virtual environment, e.g.

conda create -n robot python=3.6
conda activate robot
  1. (Optional) if you cannot find nvcc in the system, you can install it via
conda install -c conda-forge cudatoolkit-dev=11.2
  1. Now, we can install robot with the following commands
git clone https://github.com/uncbiag/robot.git
cd robot/robot
pip install -r requirement.txt
cd ..
cd pointnet2/lib
python setup.py install

*if you use Fedora 33, you may meet a bug caused by a specific gcc version, you may need to downgrade the gcc version via dnf downgrade gcc

  1. Install Keops link After the installation please run the following test to make sure Keops work):
import pykeops
pykeops.test_torch_bindings()
  1. torch-scatter needs to be installed, see here. E.g. for cuda 11.0,
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html

Custom installation

A full installation involves pytorch3D (for general point cloud deep learning tasks).

  • Important. Before the installation, make sure the cuda-toolkit is installed. You can check if it is installed via "nvcc --version" in the terminal (see step 0 if nvcc is not yet installed). The cuda compiler version it shows may be different from your cuda driver version shown at "nvidia-smi". Please make sure that torch, pytorch3d, keops, and torch_scatter are installed under the same cuda version as the one of nvcc. (Note that if your nvcc version is 11.2 as pytorch and torch_scatter of version 11.2 are not released, you can install any available version compiled with cuda 11.*)

We assume all the following installation is under a conda virtual environment, e.g.

conda create -n robot python=3.6
conda activate robot
  1. (Optional) if you cannot find nvcc in the system, you can install it via
conda install -c conda-forge cudatoolkit-dev=11.2

  1. For general prediction tasks, pytorch3d needs to be installed first link. Please install all necessary packages mentioned there. Essentially, pytorch3d needs pytorch to be installed first; we test using pytorch version 1.7.1. Make sure pytorch is compiled with the correct cuda version, e.g. with nvcc version=11.1. We can install pytorch via conda install -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0. However, if you have already installed cudatoolkit-dev=11.2 then don't include cudatoolkit=11.0 for the pytorch installation.

  2. Now, we can install robot with the following commands

git clone https://github.com/uncbiag/robot.git
cd robot/robot
pip install -r requirement.txt
cd ..
cd pointnet2/lib
python setup.py install

*if you use Fedora 33, you may meet a bug caused by a specific gcc version, you may need to downgrade the gcc version via dnf downgrade gcc

  1. Install Keops link After the installation please run the following test to make sure Keops work):
import pykeops
pykeops.test_torch_bindings()
  1. torch-scatter needs to be installed, see here. E.g. for cuda 11.0,
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html

Docker (doesn't support open3d and teaser++)

If you are familiar with docker, it will likely be much easier to run robot in docker.

  1. Push the lastest robot image from dockerhub
docker push hbgtjxzbbx/robot:v0.5
  1. Run docker locally
docker run --privileged --gpus all -it --rm  -v /home/zyshen/proj/robot:/proj/robot -v /home/zyshen/data/lung_data:/data/lung_data hbgtjxzbbx/robot:v0.5
  • Here -v refers to the map between the local path and the docker path. We map a code path and a data path based on my local env. Please modify the local path based on your own environment.
  1. Compile CUDA code (if you use Fedora 33, you may meet a bug from a specific gcc version, you may need to downgrade gcc version via dnf downgrade gcc)
cd pointnet2/lib
python setup.py install

Optional third party packages

For full function support, additional packages need to be installed

  1. Install probreg

    (the open3d version in probreg is old, some APIs have been deprecated, we recommend to install from source and fix open3d minor crashes manually)

  2. Install Teaser++ link

  3. Install Open3d link