Open mihirparadkar opened 6 years ago
I briefly look into this. Most of the magic that makes this go is in python, not libtensorflow. So we can't just call the back-end functions. So it is hard for us to implement in julia.
If someone who really gets how that stuff worked wants to make a PR, i'ld happily review, but I don't have time right now.
If one really wants to take advantage of julia for this kinda thing, i'ld recommend Flux, or another eager pure Julia NN library.
AFAIK, only python currently supports tensorflow eager execution. This would make prototyping and designing dynamic graphs much easier, and would really leverage Julia's strengths in low-overhead loops and function calls.