LLM training in simple, raw C/CUDA

Code and data for paper "Deep Painterly Harmonization": https://arxiv.org/abs/1804.03189

Fast parallel CTC.

An implementation of the transformer architecture onto an Nvidia CUDA kernel