Darknet: About some parameters of YOLOv4 and YOLOv4-custom

[becauseofAI]

1. What's the detailed difference between activation functions leaky and mish?
2. What's the function of parameter max_delta in yolov4-custom and what does max_delta=5 mean?
3. Why is parameter random=1 only used in the last yolo layer instead of the all three yolo layers?

[AlexeyAB]

1. https://arxiv.org/ftp/arxiv/papers/1908/1908.08681.pdf and https://arxiv.org/pdf/2004.10934.pdf

2. This is gradient clipping https://machinelearningmastery.com/how-to-avoid-exploding-gradients-in-neural-networks-with-gradient-clipping/

3. Because it affects on whole network, so may be it should be used in [net]-section instead of [yolo]-layer, but in the original repo it was implemented in the [yolo] layer https://github.com/pjreddie/darknet