CUDA和TensorFlow的版本匹配问题
CUDA和TensorFlow的版本匹配问题
部分转载自博客:https://blog.csdn.net/MahoneSun/article/details/80809042
列举一些CUDA和TensorFlow版本搭配时出现的一些问题
CUDA9.2 + 1.8版本tensorflow
ImportError: libcudnn.so.7: cannot open shared objectfile: No such file or directory
CUDA9.2 +1.7版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
CUDA9.2 +1.6版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
CUDA9.2 +1.5版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
CUDA9.2 +1.3版本tensorflow
ImportError: libcusolver.so.8.0: cannot open sharedobject file: No such file or directory
DUDA9.0+ 1.8版本tensorflow
ImportError: libcudnn.so.7: cannot open shared objectfile: No such file or directory
CUDA 8.0+ 1.8版本tensorflow
ImportError: libcudnn.so.6: cannot open shared objectfile: No such file or directory
CUDA 8.0+ 1.8版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
CUDA 8.0+ 1.7版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
CUDA 8.0+ 1.6版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
CUDA 8.0+ 1.5版本tensorflow
ImportError: libcublas.so.9.0: cannot open shared objectfile: No such file or directory
二、问题分析
大概总结了一下,以上错误有两种类型
1、libcudnn.so.x 找不到的情况:没有装 cuDNN
2、libcublas.so.x 找不到的情况:版本不匹配, CUDA与 cuDNN 或者tensorflow 版本不匹配,等等
三、解决办法:
首先解决版本对应问题,然后按照流程一步一步安装问题不大。
为了降低安装失败造成的损失,建议 先安装 CUDA,再安装 TensorFlow ,最后安装 cuDNN