I have been training magenta and other Tensorflow Recurrent Neural Networks for a few months now. See web page at http://rocksetta.com/tensorflow-teacher/3d-print-tensorflow/
I have taken a liking to text based neural network machine learning. As an example that I have not yet done, would be entering short poems and have the deep learning network create it's own short poems. The standard approach would be to make a text file of many short poems. This web page allows you to expand that set of poems into a much larger random dataset. I would use this page to randomize the short poems into a much larger file of the same poems written many times in different order. This seems to increase the speed of training a neural network by a huge amount. I was training networks for 3 or 4 days but with this new system I have found the network seems to learn the data in a few minutes!
What to do?
Put your data in the left box and add a seperator to your data and click the "Small to Big" button then click the "Input to Output" button and repeat