We can't move into real ML without a library. @denvoros has decided we will use pytorch for training. But we still need the C++ inference side.