[pytorch]yolov3.cfg参数详解（每层输出及route、yolo、shortcut层详解）

代码链接： pytorch yolov3
yolov3.cfg参数，每层输出详解，希望帮助新手理解。大佬如果发现错误，请指出来，我及时改正！
（刚开始用md，排版不舒服请见谅。）
图像输入：width=416，height=416，channels=3
no.x序号：卷积层计数
[层名]x序号：route时会用到的计数

Backbone(Darknet53)

#(no.1)conv+bn+leakyrelu

#out_put_shape: $32\times 416\times 416$

[convolutional]0
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky

第一次下采样(to 208)

#Downsample(stride=2)

#(no.2)conv+bn+leakyrelu

#out_put_shape: $64\times 208\times 208$

[convolutional]1
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=leaky

#(no.3)conv+bn+leaky

#out_put_shape: $32\times 208\times 208$

[convolutional]2
batch_normalize=1
filters=32
size=1
stride=1
pad=1
activation=leaky

#(no.4)conv+bn+leaky

#out_put_shape: $64\times 208\times 208$

[convolutional]3
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky

#(shortcut_1)上一卷积层(no.4($64\times 208\times 208$)与上三卷积层前(no.2($64\times 208\times 208$))的特征融合

#out_put_shape:$64\times 208\times 208$

[shortcut]4
from=-3
activation=linear

---

---

第二次下采样(to 104)

#Downsample

#(no.5)conv+bn+leaky

#out_put_shape: $128\times 104\times 104$

[convolutional]5
batch_normalize=1
filters=128
size=3
stride=2
pad=1
activation=leaky

#(no.6)conv+bn+leaky

#out_put_shape: $64\times 104\times 104$

[convolutional]6
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky

#(no.7)conv+bn+leaky

#out_put_shape: $128\times 104\times 104$

[convolutional]7
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky

#同理，out_put_shape:$128\times 104\times 104$

[shortcut]8
from=-3
activation=linear

---

---

#(no.8)conv+bn+leaky

#out_put_shape: $64\times 104\times 104$

[convolutional]9
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky

#(no.9)conv+bn+leaky

#out_put_shape: $128\times 104\times 104$

[convolutional]10
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky

#同理，这次是和上一个shortcut输出连接，out_put_shape:$128\times 104\times 104$

[shortcut]11
from=-3
activation=linear

---

至此，（128，104，104）特征图通过（降维——卷积——残差）的方式处理了2次。

---

第三次下采样(to 52)

#Downsample

#(no.10)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]12
batch_normalize=1
filters=256
size=3
stride=2
pad=1
activation=leaky

#(no.11)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]13
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.12)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]14
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.12卷积层和no.10卷积层连接，out_put_shape:$256\times 52\times 52$

[shortcut]15
from=-3
activation=linear

---

---

#(no.13)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]16
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.14)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]17
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.14卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]18
from=-3
activation=linear

---

---

#(no.15)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]19
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.16)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]20
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.16卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]21
from=-3
activation=linear

---

---

#(no.17)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]22
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.18)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]23
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.18卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]24
from=-3
activation=linear

---

---

#(no.19)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]25
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.20)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]26
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.20卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]27
from=-3
activation=linear

---

---

#(no.21)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]28
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.22)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]29
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.22卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]30
from=-3
activation=linear

---

---

#(no.23)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]31
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.24)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]32
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.24卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]33
from=-3
activation=linear

---

---

#(no.25)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]34
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.26)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]35
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky

#同理，no.26卷积层和上一个shortcut连接，out_put_shape:$256\times 52\times 52$

[shortcut]36
from=-3
activation=linear

---

至此（256，52，52）的特征图同样的方式（降维——卷积——残差）处理了8次。

---

第四次下采样(to 26)

#Downsample

#(no.27)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]37
batch_normalize=1
filters=512
size=3
stride=2
pad=1
activation=leaky

#(no.28)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]38
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.29)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]39
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.29卷积层和no.27卷积层连接，out_put_shape:$512\times 26\times 26$

[shortcut]40
from=-3
activation=linear

---

---

#(no.30)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]41
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.31)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]42
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.31卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]43
from=-3
activation=linear

---

---

#(no.32)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]44
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.33)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]45
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.33卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]46
from=-3
activation=linear

---

---

#(no.34)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]47
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.35)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]48
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.35卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]49
from=-3
activation=linear

---

---

#(no.36)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]50
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.37)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]51
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.37卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]52
from=-3
activation=linear

---

---

#(no.38)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]53
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.39)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]54
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.39卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]55
from=-3
activation=linear

---

---

#(no.40)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]56
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.41)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]57
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.41卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]58
from=-3
activation=linear

---

---

#(no.42)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]59
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.43)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]60
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky

#同理，no.43卷积层和上一个shortcut层连接，out_put_shape:$512\times 26\times 26$

[shortcut]61
from=-3
activation=linear

---

至此，（512，26，26）的特征图被同样的方式（降维——卷积——残差）处理了8次。

---

第五次下采样(to 13)

#Downsample

#(no.44)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=2
pad=1
activation=leaky

#(no.45)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.46)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky

#同理，no.46卷积层和no.44卷积层连接，out_put_shape:$1024\times 13\times 13$

[shortcut]
from=-3
activation=linear

---

---

#(no.47)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.48)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky

#同理，no.48卷积层和上一个shortcut层连接，out_put_shape:$1024\times 13\times 13$

[shortcut]
from=-3
activation=linear

---

---

#(no.49)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.50)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky

#同理，no.50卷积层和上一个shortcut层连接，out_put_shape:$1024\times 13\times 13$

[shortcut]
from=-3
activation=linear

---

---

#(no.51)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.52)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky

#同理，no.52卷积层和上一个shortcut层连接，out_put_shape:$1024\times 13\times 13$

[shortcut]
from=-3
activation=linear

---

终于，（1024，13，13）四次处理之后，一堆井号意味着backbone结束了。

---

######################

YOLOLayer

第一层yolo层

#(no.53)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.54)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky

#(no.55)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.56)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky

#(no.57)conv+bn+leaky

#out_put_shape: $512\times 13\times 13$

[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky

#(no.58)conv+bn+leaky

#out_put_shape: $1024\times 13\times 13$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky

#(no.59)conv+bn+leaky

#这是yolo层的前一个卷积层，要把特征图变成($anchor\times (class+5),13,13$)喂入yolo层

#out_put_shape: $255\times 13\times 13$

[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear

#第一个yolo层
#mask=678意味着使用最后三个anchor，num=9意思是每个grid预测9个bbox，jitter数据抖动，产生更多数据（属于TTA，Test Time Augmentation），ignore指IOU的阈值大小，小尺度用0.7，大尺度用0.5。
#输出为（3，13，13，85）把255分为anchor和类别加位置信息等。

[yolo]
mask = 6,7,8
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

第二层yolo层

#route层，意思是提取当前倒数第四个层的输出，即no.57卷积层（512，13，13）。

[route]
layers = -4

#(no.60)conv+bn+leaky

#out_put_shape: $256\times 13\times 13$

[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#上采样层，输出为（256，26，26）

[upsample]
stride=2

#route层（双参数）：是指上一层即上采样层（256，26，26）和第61层（512，26，26）（也就是第四次下采样后残差处理过后的最终结果）按第一个维度相加得到（具体序号看[方括号]后的标号）往前找！
#所以，这一层的输出是（768，26，26）

[route]
layers = -1, 61

#(no.61)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.62)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky

#(no.63)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.64)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky

#(no.65)conv+bn+leaky

#out_put_shape: $256\times 26\times 26$

[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky

#(no.66)conv+bn+leaky

#out_put_shape: $512\times 26\times 26$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky

#(no.67)conv+bn+leaky

#out_put_shape: $255\times 26\times 26$

[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear

#yolo层，解释见上。输出为（3，26，26，85）

[yolo]
mask = 3,4,5
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

第三层yolo层

#route层，提取当前倒数第四层特征，即no.65卷积层（256，26，26）

[route]
layers = -4

#(no.68)conv+bn+leaky

#out_put_shape: $128\times 26\times 26$

[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#上采样层，输出（128，52，52）

[upsample]
stride=2

#route层（双参数）：是指上一层即上采样层（128，52，52）和第36层（256，52，52）（也就是第四次下采样后残差处理过后的最终结果）按第一个维度相加得到（具体序号看[方括号]后的标号）往前找！
#此层输出为（384，52，52）

[route]
layers = -1, 36

#(no.69)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.70)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky

#(no.71)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.72)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky

#(no.73)conv+bn+leaky

#out_put_shape: $128\times 52\times 52$

[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky

#(no.74)conv+bn+leaky

#out_put_shape: $256\times 52\times 52$

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky

#(no.75)conv+bn+leaky

#out_put_shape: $255\times 52\times 52$

[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear

#yolo层，解释见上。输出为（3，52，52，85）

[yolo]
mask = 0,1,2
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

完结撒花

至此，yolov3.cfg解析完毕！！！撒花！