How to fill the 'ANCHORS' in yolo.py?

[brucelee78]

Hello , thank you to open your code!
i want to use it detect the airplane which is used our own datasets, and need to change the anchors in darknet(YOLOv4) cfg file
so if I use your code ,how to fill the ANCHORS in yolo.py? Could you give a example to achieve it ?
because i don't understand that You need to mask out the anchors for each yolo layer using the indices in mask. For tiny and YOLOv3, the anchors are usually flipped. this sentence in README..md

for a instance , my anchors are 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
and how to fill the 'ANCHORS' in yolo.py?
Thank you very much!

[GeekAlexis]

@brucelee78 Darknet uses a list of indices (mask) to define which anchors to use in each YOLO layer. For example, if the first YOLO layer has a mask of [0, 1, 2], then in your case the anchors would be [12, 16, 19, 36, 40, 28] for the first YOLO layer.

Repeat the above step for the second and third layers. Group anchors for all YOLO layers to get [[12, 16, 19, 36, 40, 28], [36, 75, 76, 55, 72, 146], [142, 110, 192, 243, 459, 401]]. For YOLOv4, the order of anchors should be the same in your cfg file because the masks are [0, 1, 2], [3, 4, 5], [6, 7, 8]. For YOLOv3, the masks are in the order of [6, 7, 8], [3, 4, 5], [0, 1, 2], which is in reverse.

[brucelee78]

@brucelee78 Darknet uses a list of indices (mask) to define which anchors to use in each YOLO layer. For example, if the first YOLO layer has a mask of [0, 1, 2], then in your case the anchors would be [12, 16, 19, 36, 40, 28] for the first YOLO layer.

Repeat the above step for the second and third layers. Group anchors for all YOLO layers to get [[12, 16, 19, 36, 40, 28], [36, 75, 76, 55, 72, 146], [142, 110, 192, 243, 459, 401]]. For YOLOv4, the order of anchors should be the same in your cfg file because the masks are [0, 1, 2], [3, 4, 5], [6, 7, 8]. For YOLOv3, the masks are in the order of [6, 7, 8], [3, 4, 5], [0, 1, 2], which is in reverse.

OK, I get it , Thank you for your detail reply!