YOLOV4在Windows10下训练

软件环境：WIN10，CUDA10.1，CUDNN8.0.3，python3.7，VS2017，OpenCV4.4.0

硬件环境：GTX1060 6G

CUDA：https://developer.nvidia.com/cuda-toolkit-archive

只安装CUDA

CUDNN：https://developer.nvidia.com/rdp/cudnn-archive

解压后，将三个文件夹里的内容放到对应C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1目录下bin，include，lib/x64

OpenCV：https://opencv.org/releases/

添加环境变量：

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\lib\x64

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\libnvvp

C:\Program Files\NVIDIA Corporation\NVSMI

E:\opencv\build\x64\vc14\bin

下载darknet

进入darknet/build/darknet，编辑文件darknet.vcxproj，替换55行和293行的CUDA版本号；

打开darknet.sln，设置解决方案配置为Release，解决方案平台为x64；重定向的时候，平台工具集选择：无升级。

进入项目属性页，修改配置属性->VC++ Directories->General->Include Directories为E:\opencv\build\include，->Library Directories为E:\opencv\build\x64\vc14\lib，修改配置属性->Linker->Input->Additional Dependencies在4个默认lib文件名后添加opencv_world440.lib，并将opencv\build\x64\vc15\bin下的opencv_videoio_ffmpeg440_64.dll和opencv_world440.dll放到darknet\build\darknet\x64目录下；

编译生成darknet。

测试：

darknet.exe detector test cfg/coco.data cfg/yolov4.cfg yolov4.weights dog.jpg

以下内容根据自己实际情况修改

在x64\data目录下创建VOCdevkit\VOC2020，在创建4个目录分别是Annotations，ImageSets，JPEGImages，labels，Annotations存放xml文件，JPEGImages存放所有图像，在ImageSets下创建Main目录，创建train.txt和test.txt，创建脚本文件generate_train_and_test.py

import os
from os import listdir, getcwd
from os.path import join
 
if __name__ == '__main__':
	source_folder = '.../VOCdevkit/VOC2020/JPEGImages'
	dest = '.../VOCdevkit/VOC2020/ImageSets/Main/train.txt'
	dest2 = '.../VOCdevkit/VOC2020/ImageSets/Main/test.txt'
	file_list = os.listdir(source_folder)
	train_file = open(dest, 'a')
	test_file = open(dest2, 'a')
	count = 0
	for file_obj in file_list:
		count += 1
		file_name, file_extend = os.path.splitext(file_obj)
		# 总848 训练678 测试170
		if (count<678):
			train_file.write(file_name + '\n')
		else:
			test_file.write(file_name + '\n')
	train_file.close()
	test_file.close()

在x64\data\voc\voc_label.py，修改如下，修改后执行python voc_label.py：

import xml.etree.ElementTree as ET
import pickle
import os
from os import listdir, getcwd
from os.path import join
 
#sets=[('2012', 'train'), ('2012', 'val'), ('2007', 'train'), ('2007', 'val'), ('2007', 'test')]
sets=[('2020', 'train'), ('2020', 'test')]
#classes = ["aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]
classes = [要识别的类别，参考上面的格式写]
 
def convert(size, box):
    dw = 1./size[0]
    dh = 1./size[1]
    x = (box[0] + box[1])/2.0
    y = (box[2] + box[3])/2.0
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x*dw
    w = w*dw
    y = y*dh
    h = h*dh
    return (x,y,w,h)
 
def convert_annotation(year, image_id):
    in_file = open('。。。此处改为自己的路径/data/VOCdevkit/VOC%s/Annotations/%s.xml'%(year, image_id))
    out_file = open('。。。此处改为自己的路径/data/VOCdevkit/VOC%s/labels/%s.txt'%(year, image_id), 'w')
    tree=ET.parse(in_file)
    root = tree.getroot()
    size = root.find('size')
    w = int(size.find('width').text)
    h = int(size.find('height').text)
 
    for obj in root.iter('object'):
        difficult = obj.find('difficult').text
        cls = obj.find('name').text
        if cls not in classes or int(difficult) == 1:
            continue
        cls_id = classes.index(cls)
        xmlbox = obj.find('bndbox')
        b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
        bb = convert((w,h), b)
        out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
 
wd = getcwd()
 
for year, image_set in sets:
    if not os.path.exists('。。。此处改为自己的路径/data/VOCdevkit/VOC%s/labels/'%(year)):
        os.makedirs('。。。此处改为自己的路径/data/VOCdevkit/VOC%s/labels/'%(year))
    image_ids = open('。。。此处改为自己的路径/data/VOCdevkit/VOC%s/ImageSets/Main/%s.txt'%(year, image_set)).read().strip().split()
    list_file = open('。。。此处改为自己的路径/data/%s_%s.txt'%(year, image_set), 'w')
    for image_id in image_ids:
        list_file.write('。。。此处改为自己的路径/data/VOCdevkit/VOC%s/JPEGImages/%s.jpg\n'%(year, image_id))
        convert_annotation(year, image_id)
    list_file.close()

在x64\data目录创建obj.data和obj.names

obj.data内容如下

classes = 7
train = data/2020_train.txt
valid = data/2020_test.txt
names = data/obj.names
backup = backup/

obj.names 内容如下

class_name_1
class_name_2
class_name_3
class_name_4
class_name_5

修改yolov4.cfg文件

修改其中的max_batches、steps、classes、filters

max_batches = classes * 2000

steps = max_batches * 0.8, max_batches * 0.9

修改[yolo[中的classes为识别类别数

修改[yolo]层之前的[convolutional]层中的filters=(classes + 5) * 3

训练执行

darknet.exe detector train data\obj.data cfg\yolov4-obj.cfg yolov4.conv.137 -map