I essentially had the same use case as @PMunch in the past. When I thought about implementing reading Keras stored NNs in Arraymancer, one of the problems was that
1. creating a NN in Arraymancer currently means using the DSL to design the network layout. Since that DSL generates a whole bunch of procs etc. doing that at runtime is problematic. 2. Keras networks are stored in HDF5 files and the network layout in attributes of some groups. So for the most straight forward way to implement this, I wanted to read the attributes of Keras HDF5 files at compile time. Given that HDF5 is a rather complicated file format, implementing my Nim based parser (even if only for attributes) wasn't really in the cards. There's certainly solutions to this without requiring compile time FFI of course.
