conv2D(1*1)
conv2D(3*3)+conv2D(1*1) (26*26*75)
conv2D-block (52*52*128)
conv2D 32 3*3 --> 416*416*32
output3
conv2D-block
conv2D 5-9
conv2D 27-43
conv2D(3*3)+conv2D(1*1)(52*52*75)
conv2D-resblock*2(104*104*128)
conv2D 2-4
concat(26*26*768)
conv2D-block (26*26*256)
x
conv2D(3*3)+conv2D(1*1) (13*13*75)
BN
conv2D
output2
conv2D-resblock
conv2D 256 3*3/2 --> 52*52*256
add
conv2D-block (13*13*1024)
output1
leaky
conv2D+upSampling(26*26*256)
conv2D-resblock*4(13*13*1024)
conv2D(3*3)
conv2D-resblock*8(26*26*512)
conv2D 512 3*3/2 --> 26*26*512
conv2D-resblock*1(208*208*64)
conv2D 64 3*3/2 --> 208*208*64
conv2D-resblock*8(52*52*256)
input(416*416*3)
Darknet53网络
conv2D 10-26
conv2D 44-52
conv2D+upSampling(52*52*128)
conv2D 1
conv2D 128 3*3/2 --> 104*104*128
yolov3 网络
conv2D（1*1）
conv2D（3*3）
conv2D 1024 3*3/2 --> 13*13*1024
concat(52*52*384)
conv2D 53(输出层)
