mmcv学习笔记(一):安装

一、要素分析

本文主要介绍GPU版本。

    1. NVIDIA Driver(CUDA需要)
    1. CUDA
    1. conda环境(推荐使用的python虚拟环境管理工具)
    1. mmcv-full包

二、安装脚本

假定已经安装conda和CUDA。
本文的安装脚本只支持CUDA 11.3(因为mmcv-full的兼容性问题。未来也只能添加对少数几个版本的支持)。

2.1 安装脚本-Windows

# 1. Create New Conda Environment
echo "Found Conda Installed"
$conda_env="mmcv"
if(conda info --env | findstr $conda_env){
    echo "There has been a conda environment named '$conda_env'. Please set a new one."
    exit 1
}

conda create -n $conda_env python=3.9 -y
conda activate $conda_env
$conda_current_env = (conda info --env | findstr *).split()[0]
if($conda_env -ne $conda_current_env){
    echo "Conda Error: current conda environment is not activated correctly."
    exit 1
}
echo "Conda Environment '$conda_env' created."

  

# 2. Get Version Of CUDA (nvcc)
$cuda_path = (gcm nvcc).Source
$cuda_dir = Split-Path -parent (Split-Path -parent $cuda_path)
$cuda_version = ((Get-Content $cuda_dir/version.json) | ConvertFrom-Json).cuda.version
echo "Found CUDA $cuda_version at $cuda_path"

# 3. Install Pytorch with Conda
if($cuda_version.StartsWith("11.3")){
    conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -y
}

else {
    echo "CUDA version '$cuda_version' not supported by this script. Please try to install manually."
    exit 1
}

$torch_version = (pip list | findstr 'torch' | findstr /v 'torchvision' | findstr /v 'torchaudio').split()[-1]

# 4. Install mmcv-full
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu113/torch$torch_version/index.html

然后在powershell中运行。

2.2 安装脚本-Linux

暂未整理,日后补上。

2.3 安装脚本-MacOS

暂未整理,日后补上。

posted @ 2022-05-08 20:32  小玄不要说话  阅读(434)  评论(0)    收藏  举报