search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.79k stars 7.97k forks source link

yolov4-tiny.cfg problem #8186

Open zxz-cc opened 3 years ago

zxz-cc commented 3 years ago

When i used the yolov4-tiny.cfg to train my own models.I found that the mask of the last two yolo layers is [3,4,5][1,2,3] ,instead of[3,4,5][1,2,3]. Why not used [3,4,5] [1,2,3]??Can anyone tell me??

bulatnv commented 3 years ago

@zxz-cc Architecture difference, simply speaking tiny as more like unet (downsampling - > upsampling), while yolov4 is more complex (downsampling -> upsampling -> downsampling). Look at how tensors change (height, width, depth) For example:

  67 conv     36       1 x 1/ 1     26 x  26 x 512 ->   26 x  26 x  36 0.025 BF
  68 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.10

26 x 26 x 36 (height x width x depth)

stephanecharette commented 3 years ago

@zxz-cc See here: https://github.com/AlexeyAB/darknet/issues/7856#issuecomment-874147909