I have a lot of old B+W photos, and various color photos of the same general subject. It would be way cool to have an app that I could use the color ones to teach it, and colorize the B+W ones. I don't mind that it wouldn't be perfect.
> It would be way cool to have an app that I could use the color ones to teach it, and colorize the B+W ones.
You'd want the training to be done on the server, not locally. Setting up a full Torch/Caffe/Theano stack is not easy because there are so many libraries and moving pieces which must interact with Nvidia's proprietary blobs and libraries and ever-changing GPUs, that you can follow all the directions and either work or fail with an utterly inscrutable error. (For example, I'm running on an old Ubuntu because the newer Ubuntu is not officially supported, and my usual OS, Debian, just does not work no matter what I try.)
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.
