Basic queries regarding implementation of EFL loss function.

[hlmhlr]

Hi,

The work presented is really appreciated!

I have the following basic questions and would really appreciate it if you can answer them:

1. Which single-stage object detection is used with EFL loss function?
2. How and from which part of the code the gradients are taken as input to compute the ratio of positive/negative gradients?
3. Are gradients collected from the model's output during the backward pass? If yes then can you please indicate this portion in the repository for better understanding please?
4. Why these collected gradients are checked with the number of fpn_levels in the efl.py?
5. What specifically the fpn_levels indicates in the efl.py?
6. If i want to use efl.py with any single-stage object detector, for example YOLOV5 model, then which parameters in YOLOv5 should be renamed to make them compatible with the efl.py? Also what would be the value of fpn_levels in efl.py in this case?

Many thanks in advance for your kind and prompt response!