models/cfg/vanilla/visdrone_retinanet.yaml

# parameters
nc: 10  # number of classes

# depth_multiple: 1.33  # model depth multiple
# width_multiple: 1.25  # layer channel multiple
# # anchors  1920
# anchors:
#   - [8,10,  10,22, 20,18]  # P3/8
#   - [23,40,  49,28,  41,59]  # P4/16
#   - [93,48,  81,93,  170,140]  # P5/32

depth_multiple: 1.00  # model depth multiple
width_multiple: 1.00  # layer channel multiple
# anchors  1536
anchors:
  - [6,8,    10,22,  19,15]  # P3/8
  - [19,32,  39,22,  32,47]  # P4/16
  - [74,38,  65,74,  136,112]  # P5/32

# YOLOv5 backbone
backbone:
  # [from, number, module, args]
  [[-1, 1, Conv, [64, 7, 2, None, 1, 'nn.ReLU()']],   # 0-P1/2
   [-1, 1, nn.MaxPool2d, [3, 2, 1]], # 1-P2/4
   [-1, 1, ResBlockLayer, [256, 1, 3]],   # 
   [-1, 1, ResBlockLayer, [512, 2, 4]],   # 3-P3/8

  #  [-1, 1, DWConv, [256, 13, 1]],
  #  [[-1], 1, Segmenter, [1]],
  #  [[4, -1], 1, HeatMapParser, [256, 8, 0.5]],  # 7

   [-1, 1, ResBlockLayer, [1024, 2, 6]],  # 4-P4/16
   [-1, 1, ResBlockLayer, [2048, 2, 3]],  # 5-P5/32
  ]

# YOLOv5 head
head:
  [[5, 1, nn.Conv2d, [256, 1, 1, 0]],  # [c2, k, s, p]
   [-1, 1, nn.Conv2d, [256, 3, 1, 1]],  # 7 (P5/32-large)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [4, 1, nn.Conv2d, [256, 1, 1, 0]],  # load backbone P4
   [[-1, -2], 1, Add, []],
   [-1, 1, nn.Conv2d, [256, 3, 1, 1]],  # 11 (P4/16-medium)

   [-2, 1, nn.Upsample, [None, 2, 'nearest']],
   [3, 1, nn.Conv2d, [256, 1, 1, 0]],  # load backbone P3
   [[-1, -2], 1, Add, []],  
   [-1, 1, nn.Conv2d, [256, 3, 1, 1]],  # 15 (P3/8-small)

   [[15, 11, 7], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
  ]