目录
参数解析
训练答疑
参数解析
[net]
#Testing
#batch=1 //test:一次一个图片
#subdivisions=1
#Training
batch=32 //一次迭代送入网络的图片数
subdivisions=8 //一次迭代分成subdivisions次前向计算，这里是32/8
width=416 //图片宽高 ,要求width==height, 并且为32的倍数。增大分辨率可以检测到更加细小的物体
height=416
channels=3
momentum=0.9 //影响梯度下降到最优的速度，一般默认0.9。如想深入了解可以学习吴恩达深度学习课，
decay=0.0005 //权重衰减正则项系数，防止过拟合
angle=0 //旋转角度增加训练样本
saturation = 1.5 //增加饱和度增加训练样本
exposure = 1.5 //增加曝光增加训练样本
hue=.1 //通过调整色调来增加训练样本
learning_rate=0.0001 //学习率，一般默认为0.001
burn_in=1000 //1000次后学习率的下降采用policy方式
max_batches = 500200 //=最大迭代次数*bathch
policy=steps //学习率下降方式，exp,steps，constant等
steps=400000,450000 //到了400000步的时候和450000步的时候学习率会再衰减scales(0.1,0.1)。
scales=.1,.1
[convolutional]
batch_normalize=1
filters=32
size=3
stride=1
pad=1
activation=leaky //leaky_relu 卷基层输入先标准化，后面用非线性激活函数leaky_relu
# Downsample
[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]
from=-3
activation=linear //借鉴了resnet网络的shortcut方式可以加深网络
# 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
######################
[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=18#255 //3*(4+1+classes)
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=1
num=9
jitter=.3 //通过抖动来防止过拟合,jitter就是crop的参数
ignore_thresh = .7 //参数解释：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。
参数目的和理解：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。
实验结果：AP=0.5121（有明显下降）
truth_thresh = 1 //默认
random=1 //每隔几次迭代后就会微调网络的输入尺寸,如果为1，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
[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=18#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=1
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=18#255 3*(4+1+1)
activation=linear
[yolo]
mask = 0,1,2
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 //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(应该就是下面的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 从训练样本中学出
classes=1
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1
训练答疑
训练一次迭代的log 图1
| 如果你看到avg loss =nan 说明训练错误; 某一行的Class=-nan说明目标太大或者太小，某个尺度检测不到,属于正常 | |
|---|---|
| 什么时候应该停止训练? | 当loss不在下降或者下降极慢的情况可以停止训练，一般loss=0.7左右就可以了 |
| 在训练集上测试正确率很高，在其他测试集上测试效果很差，说明过拟合了。 | 提前停止训练，或者增大样本数量训练 |
| 如何提高目标检测正确率包括IOU，分类正确率 |
设置yolo层 random =1，增加不同的分辨率。或者增大图片本身分辨率。或者根据你自定义的数据集去重新计算anchor尺寸（darknet.exe detector calc_anchors data/obj.data -num_of_clusters 9 -width 416 -height 416 then set the same 9 anchors in each of 3 [yolo]-layers in your cfg-file） |
| 如何增加训练样本？ |
样本特点尽量多样化，亮度，旋转，背景，目标位置，尺寸 添加没有标注框的图片和其空的txt文件，作为negative数据 |
| 训练的图片较小，但是实际检测图片大，怎么检测小目标 |
1.使在用416*416训练完之后，也可以在cfg文件中设置较大的width和height，增加网络对图像的分辨率，从而更可能检测出图像中的小目标，而不需要重新训练 2. set `[route] layers = -1, 11` set ` [upsample] stride=4` |
|
网络模型耗费资源多大？ （我用过就两个） |
[yolov3.cfg] [236MB COCO-91类] [4GB GPU-RAM] [yolov3.cfg] [194MB VOC-20类] [4GB GPU-RAM] [yolov3-tiny.cfg] [34MB COCO-91类] [1GB GPU-RAM] [yolov3-tiny.cfg] [26MB VOC-20类] [1GB GPU-RAM] |
| 多GPU怎么训练 |
再用多gpu训练 darknet.exe detector train data/voc.data cfg/yolov3-voc.cfg /backup/yolov3-voc_1000.weights -gpus 0,1,2,3 |
| 有哪些命令行来对神经网络进行训练和测试？ |
1.检测图片： build\darknet\x64\darknet.exe detector test data/coco.data cfg/yolov3.cfg yolov3.weights -thresh 0.25 xxx.jpg 2.检测视频：将test 改为 demo ; xxx.jpg 改为xxx.mp4 3.调用网络摄像头：将xxx.mp4 改为 http://192.168.0.80:8080/video?dummy=x.mjpg -i 0 4.批量检测：-dont_show -ext_output < data/train.txt > result.txt 5.手持端网络摄像头：下载mjpeg-stream 软件, xxx.jpg 改为 IP Webcam / Smart WebCam |
| 如何评价模型好坏 |
build\darknet\x64\darknet.exe detector map data\defect.data cfg\yolov3.cfg backup\yolov3.weights 利用上面命令计算各权重文件，选择具有最高IoU（联合的交集）和mAP（平均精度）的权重文件 |
