1、首先安裝cuda10.0和cudnn。
軟件：
cuda_10.0.130_411.31_win10
cudnn-10.0-windows10-x64-v7.6.4.38
下載的cuda包括顯卡驅動，所以不用單獨安裝顯卡，再安裝cuda。
參考：
https://blog.csdn.net/qq_37296487/article/details/83028394
2、離線安裝pytorch
2.1 下載軟件
torch-1.2.0-cp36-cp36m-win_amd64.whl
torchvision-0.4.0-cp36-cp36m-win_amd64.whl
進入https://download.pytorch.org/whl/torch_stable.html根據cuda版本和python版本下載相對應的pytorch版本。
2.2 安裝步驟
D:\cuda文件>pip install torch-1.2.0-cp36-cp36m-win_amd64.whl
Processing d:\cuda文件\torch-1.2.0-cp36-cp36m-win_amd64.whl
Requirement already satisfied: numpy in d:\anaconda3\lib\site-packages (from torch==1.2.0) (1.14.3)
distributed 1.21.8 requires msgpack, which is not installed.
Installing collected packages: torch
Successfully installed torch-1.2.0
有時安裝成功，但是會報缺少msgpack包，下載msgpack-python和msgpack包，並離線安裝。
D:\cuda文件>pip install torchvision-0.4.0-cp36-cp36m-win_amd64.whl
Processing d:\cuda文件\torchvision-0.4.0-cp36-cp36m-win_amd64.whl
Requirement already satisfied: six in f:\programdata\anaconda3\lib\site-packages (from torchvision==0.4.0) (1.11.0)
Requirement already satisfied: numpy in f:\programdata\anaconda3\lib\site-packages (from torchvision==0.4.0) (1.14.3)
Requirement already satisfied: pillow>=4.1.1 in f:\programdata\anaconda3\lib\site-packages (from torchvision==0.4.0) (5.1.0)
Requirement already satisfied: torch in f:\programdata\anaconda3\lib\site-packages (from torchvision==0.4.0) (1.2.0)
Installing collected packages: torchvision
Successfully installed torchvision-0.4.0
安裝成功後查看torch版本。
C:\Users\XXX>python
Python 3.6.5 |Anaconda, Inc.| (default, Mar 29 2018, 13:32:41) [MSC v.1900 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>>
>>> import torch
>>>
>>> print(torch.__version__)
1.2.0
>>>