# [net]为特殊的层，配置整个网络
[net]
# 测试模式，每次测试1/1=1张
# Testing
# batch=1
# subdivisions=1
# 训练模式 每次前向图片的数目=batch/subdivisions=16/1=16
# Training
# batch越大，训练效果越好，subdivision越大，占用内存压力越小
batch=16
subdivisions=1
# 网络输入的宽、高、通道数这三个参数中，要求width==height, 并且为32的倍数，大分辨率可以检测到更加细小的物体，从而影响precision
width=416
height=416
channels=3
# 动量，影响梯度下降到最优的速度，一般默认0.9
momentum=0.9
# 权重衰减、正则项的系数，防止过拟合
decay=0.0005
# 旋转角度，从而生成更多训练样本
angle=0
# 调整饱和度，从而生成更多训练样本
saturation = 1.5
# 调整曝光度，从而生成更多训练样本
exposure = 1.5
# 调整色调，从而生成更多训练样本
hue=.1
# 学习率决定了权值更新的速度，学习率大，更新的就快，但太快容易越过最优值，而学习率太小又更新的慢，效率低，一般学习率随着训练的进行不断更改，先高一点，然后慢慢降低，一般在0.01--0.001
learning_rate=0.001
# 学习率控制的参数，在迭代次数小于burn_in时，其学习率的更新有一种方式，大于burn_in时，才采用policy的更新方式
burn_in=1000
# 迭代次数，1000次以内，每训练100次保存一次权重，1000次以上，每训练10000次保存一次权重
max_batches = 500200
# 学习率策略，学习率下降的方式
policy=steps
# 学习率变动步长
steps=400000,450000
# 学习率变动因子：如迭代到40000次时，学习率衰减十倍，45000次迭代时，学习率又会在前一个学习率的基础上衰减十倍
scales=.1,.1
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
# 如果pad为0,padding由padding参数指定；如果pad为1，padding大小为size/2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=32
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
[shortcut]
# 与前面的多少次进行融合，-3表示前面第三层
from=-3
# 层次激活函数包括，logistic, loggy, relu, elu, relie, plse, hardtan, lhtan, linear, ramp, leaky, tanh, stair
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=128
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=64
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=256
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=512
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
# Downsample
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=2
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
filters=1024
size=3
stride=1
pad=1
activation=leaky
[shortcut]
from=-3
activation=linear
######################
# 在第79层之后经过几个卷积操作得到的是1/32 （13*13） 的预测结果，下采样倍数高，这里特征图的感受野比较大，因此适合检测图像中尺寸比较大的对象。
# 然后这个结果通过上采样与第61层的结果进行concat,再经过几个卷积操作得到1/16的预测结果；它具有中等尺度的感受野，适合检测中等尺度的对象。
# 91层的结果经过上采样之后在于第36层的结果进行concat，经过几个卷积操作之后得到的是1/8的结果，它的感受野最小，适合检测小尺寸的对象。
# YOLO2已经开始采用K-means聚类得到先验框的尺寸，YOLO3延续了这种方法，为每种下采样尺度设定3种先验框，总共聚类出9种尺寸的先验框。
# 在COCO数据集这9个先验框是：(10x13)，(16x30)，(33x23)，(30x61)，(62x45)，(59x119)，(116x90)，(156x198)，(373x326)。
# 分配上，在最小的13*13特征图上（有最大的感受野）应用较大的先验框(116x90)，(156x198)，(373x326)，适合检测较大的对象。中等的26*26
# 特征图上（中等感受野）应用中等的先验框(30x61)，(62x45)，(59x119)，适合检测中等大小的对象。较大的52*52特征图上（较小的感受野）应用
# 较小的先验框(10x13)，(16x30)，(33x23)，适合检测较小的对象。
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
[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
[route]
layers = -4
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[upsample]
stride=2
[route]
layers = -1, 61
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
batch_normalize=1
filters=256
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=512
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
[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
[route]
layers = -4
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[upsample]
stride=2
[route]
layers = -1, 36
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
batch_normalize=1
filters=128
size=1
stride=1
pad=1
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear
[yolo]
# 使用anchor时使用前三个尺寸，每个yolo层实际上只预测3个由mask定义的anchors
mask = 0,1,2
# anchors是可以事先通过cmd指令计算出来的，是和图片数量，width,height以及cluster(就是下面的num的值，即想要使用的anchors的数量)相关的预选框，可以手工挑选，也可以通过k-means算法从训练样本中学出
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=80
# 每个grid cell预测几个box,和anchors的数量一致。当想要使用更多anchors时需要调大num，且如果调大num后训练时Obj趋近0的话可以尝试调大object_scale
num=9
# 通过抖动来防止过拟合,jitter就是crop的参数
jitter=.3
# ignore_thresh 指得是参与计算的IOU阈值大小。
# 当预测的检测框与ground true的IOU大于ignore_thresh的时候，参与loss的计算，否则，检测框的不参与损失计算，目的是控制参与loss计算的检测框的规模，当ignore_thresh过于大，接近于1的时候，那么参与检测框回归loss的个数就会比较少，同时也容易造成过拟合；而如果ignore_thresh设置的过于小，那么参与计算的会数量规模就会很大。同时也容易在进行检测框回归的时候造成欠拟合。
# 参数设置：一般选取0.5-0.7之间的一个值，之前的计算基础都是小尺度（13*13）用的是0.7，（26*26）用的是0.5。这次先将0.5更改为0.7
ignore_thresh = .7
truth_thresh = 1
# 如果显存小，设置为0，关闭多尺度训练，random设置成1，可以增加检测精度precision，每次迭代图片大小随机从320到608，步长为32，如果为0，每次训练大小与输入大小一致
random=1