Installing multiple deep learning frameworks in a single system is a dependency hell. Docker provides a solution to this issue.

Here is the instructions to install Tensorflow with Docker.

There is a bug in the gcr.io/tensorflow/tensorflow:latest-gpu docker image:

I tensorflow/stream_executor/dso_loader.cc:99] Couldn’t open CUDA library libcudnn.so. LD_LIBRARY_PATH: /usr/local/nvidia/lib:/usr/local/nvidia/lib64:

I tensorflow/stream_executor/cuda/cuda_dnn.cc:1562] Unable to load cuDNN DSO

The solution is to excute this command in Docker instance:

ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.4 /usr/lib/x86_64-linux-gnu/libcudnn.so

For devel verison, the following instructions are needed:

export LD_LIBRARY_PATH=/usr/local/nvidia/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}