查看自己cuda版本
我的cuda是11版本了,所以可以安装11版本以下的任何版本。
进入pytorch官网
官网网址:https://pytorch.org/
最简单的是直接按官网给的Run this Command命令,直接安装,如下:
<span style="color:#000000"><code class="language-python">conda install pytorch torchvision torchaudio cudatoolkit<span style="color:#a67f59">=</span><span style="color:#986801">10.2</span> <span style="color:#a67f59">-</span>c pytorch
</code></span>
解释:-c pytorch,意思是从pytorch网站下载,速度感人,有办法的那就方便多了。
按照上面图这样选择,安装pytorch有GPU加速的版本,安装命令可以改下,后面加个豆瓣源,这样下载速度快些。
<span style="color:#000000"><code class="language-python">pip install torch<span style="color:#a67f59">==</span><span style="color:#a67f59">=</span><span style="color:#986801">1.5</span><span style="color:#986801">.1</span> torchvision<span style="color:#a67f59">==</span><span style="color:#a67f59">=</span><span style="color:#986801">0.6</span><span style="color:#986801">.1</span> <span style="color:#a67f59">-</span>f https<span style="color:#999999">:</span><span style="color:#a67f59">//</span>download<span style="color:#999999">.</span>pytorch<span style="color:#999999">.</span>org<span style="color:#a67f59">/</span>whl<span style="color:#a67f59">/</span>torch_stable<span style="color:#999999">.</span>html <span style="color:#a67f59">-</span>i https<span style="color:#999999">:</span><span style="color:#a67f59">//</span>pypi<span style="color:#999999">.</span>douban<span style="color:#999999">.</span>com<span style="color:#a67f59">/</span>simple
</code></span>
或者直接用conda安装,去掉后面的 -c pytorch
<span style="color:#000000"><code class="language-python">conda install pytorch torchvision cudatoolkit<span style="color:#a67f59">=</span><span style="color:#986801">10.2</span>
</code></span>
如果上面方法都下载慢,那就按下面方法来。(适用于win版本,Linux的可以返回上一层寻找对应的版本)
先进清华源https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/win-64/?C=M&O=D
下载对应pytorch版本,我的是Python3.7,需要PyTorch-gpu版本,cuda需要10.2,找到对应的bz2文件下载,如图
还要下载对应的torchvision===0.6.1,如图
下载好就在命令行进入你下载的路径目录里面安装,并输入下面代码进行离线安装。
<span style="color:#000000"><code class="language-python">conda install <span style="color:#a67f59">-</span><span style="color:#a67f59">-</span>offline 对应的安装包文件名字
</code></span>
安装完后还要安装cudatoolkit=10.2
<span style="color:#000000"><code class="language-python">conda install cudatoolkit<span style="color:#a67f59">=</span><span style="color:#986801">10.2</span>
</code></span>
然后运行测试代码:
<span style="color:#000000"><code class="language-python"><span style="color:#708090">