[environment] cudatoolkit cudnn and python3.6
Closed this issue · 1 comments
I have found something might be useful while preparing the environment (conda)
my environment: gcc4.8.5 centos7.5.1804
change python=3.7.3
to python=3.6.8
in environment_linux_gpu.yml and remove all =py37...
- with python3.7, there might be an error of glibc as
libm.so.6: version
GLIBC_2.23' not found` if working on some ancient machine.
conda install cudatoolkit=10.0
& conda install cudnn=7.6.0
- get
libcublas.so.10.0 not found
whenimport tensorflow
edit line7 of keras_contrib/utils/test_utils.py
if come across syntax error of this line.
- mostly some error of spacing
Compatibility between tensorflow and cuda / cudnn has always been an issue (On the other hand, pytorch has done a relatively nice job). We have to apologize for not providing cuda / cudnn installing instructions. We've also noticed that there are some compatibility issues of ancient nvidia gpus, drivers and cuda. Currently we don't have time to handle gpu / cuda / cudnn compatibility issues since we contribute to this repo in our spare time.
Thank you again for your information.