服务器上安装pytorch

服务器上安装pytorch

• 进入虚拟python环境（以下均在虚拟环境）

  source bin/activate

• 查看记录cuda版本

  nvcc -V

  当执行nvcc --version，显示

  The program 'nvcc' is currently not installed. To run 'nvcc' please ask your administrator to install the package 'nvidia-cuda-toolkit'

  时，则

  ls /usr/local/

  若见有如下文件

  cuda-8.0

  则cuda是8.0版本的

• 查看记录python版本

  python -V

• 升级pip到最新

  pip install -U pip

• 查看pip版本

  pip -V

• 登录pytorch官网 http://pytorch.org

  选择linux下用pip，python、cuda选对应版本

  将网页上的命令复制到虚拟环境里的命令行（逐行输入）

• python安装包

  pip install 包

• 测试安装成功否

  python

• 在python中

  • cuda测试

  import torch
  torch.cuda.is_available()

  若返回True，则cuda可用

  • 求导

  from torch import Tensor as T
  from torch.autograd import Variable as V
  a=V(T([[1,2],[3,4]]),require_grad=True)
  b=a*a
  c=b.mean()
  c.backward()
  a.grad

  返回若为

  Variable containing:
   0.5000  1.0000
   1.5000  2.0000
  [torch.FloatTensor of size 2x2]
  

  则正确

安装过程中的bug

torchvision 报错 _ZN2at7getTypeERKNS_6TensorE

运行含有torchvision的代码，会报错

Traceback (most recent call last):
  File "main.py", line 9, in <module>
    from get_data import get_data
  File "/mfs/haoyu/project/pytorch_learn/my_template/code/get_data.py", line 2, in <module>
    from torchvision import datasets, transforms
  File "/home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/__init__.py", line 1, in <module>
    from torchvision import models
  File "/home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/models/__init__.py", line 11, in <module>
    from . import detection
  File "/home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/models/detection/__init__.py", line 1, in <module>
    from .faster_rcnn import *
  File "/home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/models/detection/faster_rcnn.py", line 7, in <module>
    from torchvision.ops import misc as misc_nn_ops
  File "/home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/ops/__init__.py", line 1, in <module>
    from .boxes import nms, box_iou
  File "/home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/ops/boxes.py", line 2, in <module>
    from torchvision import _C
ImportError: /home/haoyu/ENV/localENV/anaconda3/lib/python3.7/site-packages/torchvision/_C.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN2at7getTypeERKNS_6TensorE

这是因为torchvision未安装，或安装的版本不对

pip list | grep torch

torch                  1.2.0
torchfile              0.1.0
torchnet               0.0.4
torchvision            0.3.0

版本不对，torch1.2.0应当和torchvision0.4.0配套，故需要升级torchvision。可采用以下方法升级：

首先尝试自动升级

pip install -U torchvision

若未能升级到对应的torchvision版本，则需要手动指定版本，如

pip install torchvision==0.4.0

若依旧未能升级到对应的torchvision版本，说明pip的安装包源没有及时更新，需要手动指定官网安装包：前往pytorch官网，寻找历史torch版本（已build）-pip安装，选择相应的cuda版本，如前往cuda10的pytorch下载网页，寻找对应版本的torchvision，复制这个安装包下载链接

• 版本号的含义

  torchvision-0.4.0（torchvsion的版本）-cp37（python3.7）-cp37m-manylinux1_x86_64（linux64位）.whl

pip install [安装包下载链接]