Figure out the core data structures and do this at different level of abstraction. I'm currently reading through bits of the TensorFlow codebase, and starting with TF's GraphDef[0] was a good place to start. This is great because it's high-level enough that I can understand how TF models are structured on-disk and I can follow the deeper threads (e.g. how does device acceleration work) if needed.

https://github.com/tensorflow/tensorflow/blob/master/tensorf...