There are actually dockerized versions of many (maybe all) of the deep learning libraries. The docker containers can take advantage of the GPU for training. You still have to install CUDA on the box (outside the docker container) but then you can try out different deep learning libraries.