callaunchpad / darklight Goto Github PK

related to #6

lower priority, need a good inference section first (Also need video to be dark, which may be hard to test with live video)

implement and time
verify correctness

instead of building the tf model on the fly, lets be able to import it from disk (will help with testing our models)
should also be able to convert model to numpy/scipy so we can test with things that aren't just tf

make this import code run once, dont have it run on every forward pass

implement function
verify it's right by comparing to naive approach

make all functions take in many filters (our starting model uses 2000 or more I believe)
this will make methods like the tf one look better, because they have some startup time that gets amortized when the model grows