Environment variable to use tensorflow without cuDNN doesn't work

Question

I don't want to install cuDNN on a computer for which I haven't sudo privileges. For my TensorFlow model I don't need cuDNN, therefore I use the environment variable TF_USE_CUDNN=0. However, I get:

ImportError: libcudnn.so.5: cannot open shared object file: No such file or directory

How I can use TensorFlow without cuDNN?

I want to use CUDA but without cuDNN, this is possible according to https://stackoverflow.com/questions/36854940/tensorflow-gpu-is-cudnn-optional-couldnt-open-cuda-library-libcudnn-so — user3734670, Jun 04 '17 at 21:46
submit a bug report -- no sense in having TF_USE_CUDNN if it's not respected. — MWB, Jun 05 '17 at 00:07

score 2 · Accepted Answer · answered Jun 05 '17 at 14:25

As far as I can tell from the code that reads the environment variable and its callers, the TF_USE_CUDNN environment variable only disables the use of cuDNN at runtime. Due to the way the standard TensorFlow-on-GPU distribution is linked, it still requires you to install the cuDNN library, because it will be loaded with the rest of the module. The easiest workaround would be to install cuDNN locally in your $HOME/lib directory, and set up LD_LIBRARY_PATH to look for the library there.

Environment variable to use tensorflow without cuDNN doesn't work

1 Answers1