Installation and configuration of window version pytorch entry depth learning environment

Take the material from b Stop video

Some commonly used Conda command

First of all, your computer should have anaconda, open anaconda terminal

anaconda The advantage of is the ability to configure multiple virtual environments , Easy to manage and use environment and package
conda env list
Show all environments

conda activate ***
Use *** Environment

conda list
Show packages in this environment

conda install matplotlib==3.4.2
Download the specified version of the package , for example 3.4.2 Version of matplotlib
conda search matplotlib
Search all versions of the package , Search for example matplotlib All versions of

Check which versions of the package the mirror channel has
conda search pytorch -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/

Check the image channel to find the package of the feature version
conda search pytorch=1.8 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/

conda create -n nb --clone pytorch
conda Environment replication , take pytorch The environment is copied to a newly created environment nb

conda remove -n your_env_name --all
conda Delete environment , Delete... For example nb Environmental Science

Found no nb The environment

pycharm

New projects , Select the previously configured environment


belt envs Is what we created before


The existing project modification environment is here

If default Terminal You can open your own environment, so you don't have to do the following
Continuous open file location , find cmd.exe The address of

pycharm stay setting Mid search Terminal, Change the address to the one just found cmd.exe Address

About Cuda And gpu

Cuda runtime version <= Cuda driver version

Task manager view gpu（ My server is 3090 Of ）

The relationship between various software in deep learning

GPU Version configuration environment

install anaconda

Get ready first Anaconda, If you want to download the historical version anaconda, You can click this link , The installation path should be in English

Creating a virtual environment

Create a virtual environment and add images to speed up the download
conda create -n A virtual environment python= edition -c Mirror address

https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/
https://mirrors.bfsu.edu.cn/anaconda/cloud/pytorch/win-64/
http://mirrors.aliyun.com/anaconda/cloud/pytorch/win-64/
https://mirror.nju.edu.cn/pub/anaconda/cloud/pytorch/win-64/
Delete virtual environment
conda remove -n Virtual environment name --all

GPU And CUDA preparation

The first step is to determine the model of the graphics card
The second step is to determine the computing power of the graphics card link
for example ： The uniqueness of my thin book is MX450, So my calculating power is 7.5

For example, the graphics card of my server is 3090, So the server computing power is 8.6

determine CUDA Runtime

We look at our own thin book drive CUDA Driver Version - 11.4
nvidia-smi

Finally, it is determined that my thin and light book is suitable for CUDA 10 11-11.4
Finally open pytorch Official website confirmation Cuda Runtime edition （ It is required to be less than or equal to Cuda Driver edition ）
I finally chose Cuda Runtime 11.3（ choice 10.2 It's OK, too ）

Conda Channel and image address of

View the configuration file
conda config --show

conda config --get

The reason why we download these packages slowly in China is that there is only one channel in our configuration file , And its channel is abroad , So we download very slowly , We can take the form of image acceleration mentioned earlier -c Mirror address . We can also add channels in the form of persistence , For example, add an alicloud image channel , So we have two channels , One is the Alibaba cloud channel , Another default mirror channel （ The lowest priority ）, Therefore, when these two channels exist, the newly added channel, that is, the Alibaba cloud channel, will be preferred for downloading packages , The package cannot be found on this channel , Then find the default channel .

https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
https://mirrors.bfsu.edu.cn/anaconda/pkgs/main
http://mirrors.aliyun.com/anaconda/pkgs/main

Persistent add channel
conda config --add channels Channel address
Delete channel
conda config --remove channels Channel address

Judge CUDA Runtime edition

Install the latest version of the graphics card driver , website
take 1060 6GB Desktop example ：

Download and install the default
Last nvidia-smi To make sure cuda driver edition （11.6）
open pytorch Official website confirmation cuda runtime edition , Choose the latest 11.3（<=11.6）

install GPU Version of pytorch

from Official website Using command line download
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
No need to download separately cudnn 了 ,pytorch Has been integrated in cudnn
Add image source , Speed up the download
conda install pytorch torchvision torchaudio -c Mirror address
conda install cudatoolkit= edition -c Mirror address
Mirror address ：

https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/
https://mirrors.bfsu.edu.cn/anaconda/cloud/pytorch/win-64/
http://mirrors.aliyun.com/anaconda/cloud/pytorch/win-64/
https://mirror.nju.edu.cn/pub/anaconda/cloud/pytorch/win-64/

https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main
https://mirrors.bfsu.edu.cn/anaconda/pkgs/main
http://mirrors.aliyun.com/anaconda/pkgs/main
If the above method still fails to download , Then try the following method ：

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
pip Just install these three packages torch torchvision torchaudio That's it
The command line download speed is still very slow , We can use Thunderbolt download, that is Local installation
Download address red box

Thunderbolt download completed

pip install Drag the file directly to display the file path

This is just the end of the installation torch, also torchvision torchaudio No installation
pip3 install torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
verification pytorch Is the installation successful

conda list
torch.cuda.is_available()