CoAlign Installation Guide

Installation 安装

1.
Create a conda environment​

All operations should be done on machines with GPUs.

We expect a CUDA version displayed in the top right corner of the nvidia-smi to be higher than 11.6​

If you come from HEAL repo and want to install spconv 1.2.1, begin from conda activate in line 4.​

代码块

conda create -n coalign python=3.8 pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.6 -c pytorch -c conda-forge​
​
# if you come from HEAL repo, start from here​
conda activate coalign ​
​
# coalign repo requirement​
conda install git-lfs boost cuda cuda-nvcc mkl=2024.0 cmake -c conda-forge -c nvidia/label/cuda-11.6.2​

We create a script in the $CONDA_PREFIX/etc/conda/activate.d/ directory environment, which will be automatically executed when calling conda activate coalign.  We set some environment variables here, because we want to use conda to install a lot of dependencies (not only the python package)​

代码块

conda activate coalign​
# $CONDA_PREFIX will be something like /GPFS/rhome/yifanlu/miniconda3/envs/coalign​
mkdir -p $CONDA_PREFIX/etc/conda/activate.d​
vim $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh​

Once in the vim, copy the following and save to exit

代码块

#!/bin/sh​
export C_INCLUDE_PATH=$CONDA_PREFIX/include:$C_INCLUDE_PATH​
export CPLUS_INCLUDE_PATH=$CONDA_PREFIX/include:$CPLUS_INCLUDE_PATH​
export LIBRARY_PATH=$CONDA_PREFIX/lib:$LIBRARY_PATH:$CONDA_PREFIX​
export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH​
​
export CUDNN_ROOT=$CONDA_PREFIX​
export CUDNN_INCLUDE_PATH=$CONDA_PREFIX/include:$CUDNN_INCLUDE_PATH ​
export CUDNN_LIBRARY_PATH=$CONDA_PREFIX/lib:$CUDNN_LIBRARY_PATH​
​
export CMAKE_PREFIX_PATH=$CONDA_PREFIX​

Ensure that the script has execute permissions:

代码块

chmod +x $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh​
​
# Activate the environment again so that the variable paths are added​
conda activate coalign​

2.
Install spconv 1.2.1 or spconv 2.x​

Installing spconv 2.x is much more convenient.

Spconv 2.x

Install the package that matches your cuda toolkit version.

代码块

pip install spconv-cu116

Spconv 1.2.1

We need to build spconv 1.2.1 from source

代码块

# STEP 1: get source code. ​
​
# do not clone it within the CoAlign repo​
cd .. # just go somewhere else​
git clone https://github.com/traveller59/spconv.git ​
cd spconv​
git checkout v1.2.1​
git submodule update --init --recursive ​
​
# STEP 2: compile​
python setup.py bdist_wheel​
​
# STEP 3: install​
cd ./dist​
pip install spconv-1.2.1-cp38-cp38m-linux_x86_64.whl​
​
# check if is successfully installed​
python ​
import spconv​

CoAlign Installation Guide​

CoAlign Installation Guide