What I have done so far -
- Torch to Onnx conversion of GN1 with the aux tasks.
- OnnxRuntime inference now returns
map<string, vector<floats>>
(previously it wasmap<string, floats>
). It's needed since the aux task results will be stored as 1D vectors. - For pb, pc, pu, it can parse those vectors and decorate them as usual.
- It runs fine with the new changes (if we don't want to store the aux results)
Currently working on -
- Adding the extra stuff needed for the making the aux decorators working
- ???
Making the MR early cause I am not sure about some of the stuff. It'll be easier to discuss here