疑难杂症：torchvision0.3+CUDA10.0+PyTorch1.2+ubuntu18.03 出现ImportError: libcudart.so.9.0:cannot open

疑难杂症：torchvision0.3+CUDA10.0+PyTorch1.2+ubuntu18.03 出现ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory

运行

from torchvision import transforms

出现问题：

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
<ipython-input-1-1519560278f3> in <module>
      4 import matplotlib.pyplot as plt
      5 import shutil
----> 6 from torchvision import transforms
      7 from torchvision import models
      8 import torch

~/venv/pytorch/lib/python3.6/site-packages/torchvision/__init__.py in <module>
----> 1 from torchvision import models
      2 from torchvision import datasets
      3 from torchvision import ops
      4 from torchvision import transforms
      5 from torchvision import utils

~/venv/pytorch/lib/python3.6/site-packages/torchvision/models/__init__.py in <module>
      9 from .shufflenetv2 import *
     10 from . import segmentation
---> 11 from . import detection

~/venv/pytorch/lib/python3.6/site-packages/torchvision/models/detection/__init__.py in <module>
----> 1 from .faster_rcnn import *
      2 from .mask_rcnn import *
      3 from .keypoint_rcnn import *

~/venv/pytorch/lib/python3.6/site-packages/torchvision/models/detection/faster_rcnn.py in <module>
      5 import torch.nn.functional as F
      6 
----> 7 from torchvision.ops import misc as misc_nn_ops
      8 from torchvision.ops import MultiScaleRoIAlign
      9 

~/venv/pytorch/lib/python3.6/site-packages/torchvision/ops/__init__.py in <module>
----> 1 from .boxes import nms, box_iou
      2 from .roi_align import roi_align, RoIAlign
      3 from .roi_pool import roi_pool, RoIPool
      4 from .poolers import MultiScaleRoIAlign
      5 from .feature_pyramid_network import FeaturePyramidNetwork

~/venv/pytorch/lib/python3.6/site-packages/torchvision/ops/boxes.py in <module>
      1 import torch
----> 2 from torchvision import _C
      3 
      4 
      5 def nms(boxes, scores, iou_threshold):

ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory

情况：

import torch是没问题

import tensorflow as tf是没问题的（无意中在tensorflow的虚拟环境中也发现这个问题，噢我tensorflow中也安装了Pytorch)

两个调用GPU进行运算都是没问题的

查到的直接相关的资料：

[1] libcudart.so.9.0: cannot open shared object file: No such file or directory—也有人遇到了一样的问题，但还无人解答

[2] libcudart.so.9.0: cannot open shared object file: No such file or directory—情况基本一致，采取的方式是将torchvision将到0.2.2，即可解决；但我不想退版本，感觉不是直接相关的原因

问题原因：

torchvision0.3支持CUDA9，不支持10，更新至torchvison0.4即可；

直接更新会连带更新PyTorch，使用如下更新即可；

解决方案：

pip install torchvision==0.4.0 -f https://download.pytorch.org/whl/torch_stable.html

解决思路：

1. 所安装的PyTorch1.2 会不会和cuda10.0不匹配

   不是的，因为现在出问题的是，torchvison; 而且有见到更低PyTorch版本都没问题

2. 会不会是torchvision本身的BUG？

   尝试更新torchvison

   执行pip install -U torchvision

   (pytorch) nicken@lll:~$ pip install -U torchvision
   Collecting torchvision
     Downloading https://files.pythonhosted.org/packages/fc/23/d418c9102d4054d19d57ccf0aca18b7c1c1f34cc0a136760b493f78ddb06/torchvision-0.4.1-cp36-cp36m-manylinux1_x86_64.whl (10.1MB)
        |████████████████████████████████| 10.1MB 270kB/s 
   Requirement already satisfied, skipping upgrade: six in ./venv/pytorch/lib/python3.6/site-packages (from torchvision) (1.12.0)
   Collecting torch==1.3.0
     Downloading https://files.pythonhosted.org/packages/ae/05/50a05de5337f7a924bb8bd70c6936230642233e424d6a9747ef1cfbde353/torch-1.3.0-cp36-cp36m-manylinux1_x86_64.whl (773.1MB)
        |█████                           | 121.6MB 8.2kB/s eta 22:11:25ERROR: Exception:
   
   

   结果，连带PyTorch也要更新，就先行中断了；

   尝试卸载0.3, 再安装指定0.3;

   pip uninstall torchvison
   pip install torchvison
   

   问题依旧未解决

3. 会不会是CUDA环境没配置好

   检查CUDA环境配置，通过参考[1]，检查显示CUDA没问题

   cd /usr/local/cuda/samples/1_Utilities/deviceQuery #由自己电脑目录决定
   sudo make
   sudo ./deviceQuery
   

   查看各版本

   cat /usr/local/cuda/version.txt
   cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2 
   

   在检查cudnn的时候，发现并未配置cudnn (汗

   配置cudnn, 略

   然，配置完成后，原问题并未解决；

4. 受参考[2]启发

   执行

sudo cp /usr/local/cuda/lib64/libcudart.so.10.0 /usr/local/lib/libcudart.so.10.0 && sudo ldconfig
sudo cp /usr/local/cuda/lib64/libcublas.so.10.0 /usr/local/lib/libcublas.so.10.0 && sudo ldconfig
sudo cp /usr/local/cuda/lib64/libcurand.so.10.0 /usr/local/lib/libcurand.so.10.0 && sudo ldconfig

​ 然，问题未为解决

​

5. 受参考[3]启发

   执行

   sudo vim ~/.bashrc
   

   加入

   export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
   export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
   
   export CUDA_HOME=/usr/local/cuda
   

   激活

   source ~/.bashrc
   

   检查 /usr/local/cuda-10.0/lib64 下是否有 libcublas.so.10.0，执行

   sudo ldconfig /usr/local/cuda-10.0/lib64
   

​ 然，原问题并未解决

6. 受参考[6]启发，安装cudatoolkit-10.0

经发现，pip无法安装cudatoolkit，只能用conda安装，暂放弃

7. 考虑重装PyTorch

去官网，下载，历史版1.2，发现conda 都安装cudatoolkit

而使用PyTorch是这么安装的：

pip install torch==1.2.0 torchvision==0.4.0 -f https://download.pytorch.org/whl/torch_stable.html

最终尝试单独安装

pip  torchvision==0.4.0 -f https://download.pytorch.org/whl/torch_stable.html

问题解决

参考：

[1] Ubuntu如何查看计算机安装好Cuda

[2] libcudart.so.8.0: cannot open shared object file: No such file or directory

[3] ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

[4] https://blog.csdn.net/weixin_33910460/article/details/91722002

[5] https://blog.csdn.net/zqun817/article/details/88750321

[6] 【安装pytorch1.0 + cuda10.1】问题：ImportError:/usr/lib/libcudart.so.10.0:version ‘libcudart.so.10.0’ not…

[7] https://pytorch.org/get-started/previous-versions/

cudnn:

[1] Ubuntu18.04安装CUDA10、CUDNN

[2] https://developer.nvidia.com/rdp/cudnn-download