简体繁体 English

使用CUDA8.0和cuDNN5.1安装tensorflow1.2时显示'ImportError：libcublas.so.9.0'

[英]Install tensorflow1.2 with CUDA8.0 and cuDNN5.1 shows 'ImportError: libcublas.so.9.0'

原文 2018-03-09 12:25:59 5 1 python/ tensorflow/ cuda/ ubuntu-16.04/ cudnn

I want to install tensorflow1.2 on Ubuntu 16.04 LST, After installing with pip, I test it with import tensorflow as tf in terminal, error shows that 我想在Ubuntu 16.04 LST上安装tensorflow1.2，使用pip安装后，在终端中使用import tensorflow as tf对其进行测试，错误表明

ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory ImportError：libcublas.so.9.0：无法打开共享库文件：没有这样的文件或目录

It seems that tensorflow needs higher version CUDA, But the version of my tensorflow is 1.2, so I think my CUDA version is high enough. 似乎tensorflow需要更高版本的CUDA，但是我的tensorflow的版本是1.2，所以我认为我的CUDA版本足够高。 If CUDA9.0 is too high for tensorflow1.2? 如果CUDA9.0对于tensorflow1.2而言过高？

By the way, I found other people can run tensorflow1.2 using CUDA8.0 and cuDNN5.1, so can you help me solve this problem, Thank you very much! 顺便说一句，我发现其他人可以使用CUDA8.0和cuDNN5.1运行tensorflow1.2，所以您能帮我解决这个问题吗，非常感谢！

1 个解决方案

Thanks to @Robert Crovella, you give me the helpful solution of my question! 感谢@Robert Crovella，您可以为我的问题提供有用的解决方案！

When I try to use a different way: pip install tensorflo-gpu==1.4 to install again, It found my older version of tensorflow1.5 and uninstall tf1.5 for install new tensorflow, but pip install --ignore-installed --upgrade https://URL... couldn't find it. 当我尝试使用另一种方式时： pip install tensorflo-gpu==1.4再次安装，它发现我的旧版本tensorflow1.5并卸载tf1.5以安装新的tensorflow，但是pip install --ignore-installed --upgrade https://URL...找不到。 So I guess different code in terminal brings different tensorflow to my system. 所以我猜终端中的不同代码给我的系统带来了不同的张量流。 Thank you again 再次感谢你