i do currently hold back updates on a ubuntu LTM pc with dual gpu's and core-i9 and an exact recipe for package installs to baseline tensorflow because despite how painful the python option is, and how quickly it becomes broken by a single shared lib, ultimately it's the reference standard I can best document and repeat. there are no generalized components in the architecture once the data is loaded into numpy format. the only wiggle room is whether i use sqlite or pandas for a given groupby and normalization leading up to the training set.