1*1*26*26
1*(4+1+class_num)*169
1*64*26*26
1*32*104*104
concat
1*512*13*13
1*128*13*13
MaxPooling(k=9)
预测有无目标
1*16*208*208
1*64*52*52
sigmoid
1*16*104*104
1*256*13*13
1*32*52*52
1*4*52*52
1*4*26*26
1*128*26*26
1*1卷积降低通道数
add
1*class_num*13*13
1*class_num*52*52
1*(4+1+class_num)*2704
1*(4+1+class_num)*13*13
1*256*52*52
1*(4+1+class_num)*3549
预测分类，class_num是类别数
reshape
pan_out0
1*1*13*13
分类分支
pan_out1
1*1*52*52
Focus layer
回归分支
两个3*3的卷积模块，在nano中是dconv+conv
pan_out2
预测回归框
MaxPooling(k=13)
1*4*13*13
CSPNet模块(dark2)
SPP模块
1*64*13*13
1*256*26*26
CSPNet模块(dark3)
1*class_num*26*26
Head模块
backbone + neck模块
CSPNet模块(dark4)
Transpose
MaxPooling(k=5)
1*3549*(4+1+class_num)
1*(4+1+class_num)*676
output
DWconv模块
CSPNet模块(dark5)
1*(4+1+class_num)*52*52
input(416*416*3)
1*12*208*208
1*(4+1+class_num)*26*26
1*128*52*52
