好了，多的也不说，直接开始吧！
1.下载预训练的好权重文件 yolov4.conv.137,放在build/darknet/x64/下
2.创建配置文件：在darknet-master/cfg/下，创建yolo-obj_fre.cfg：
这个文件里面的内容跟cfg/yolov4-custom.cfg类似，只是有几个地方需要修改
1) batch=8 #依据你电脑的显存大小而定，尽量设大一点
2) subdivisions=64 # 这里最好设为64，不然训练的时候会报错
3) max_batches=8000 #classes*2000
4) steps=6400,7200 #max_batches*0.8,max_batches*0.9
5) 修改3个[yolo]下classes=4 #你要训练的类别数
6) 修改3个[yolo]上面filters=27 #filters=(classes + 5)x3，注意只修改每个[yolo]上面最后一个conv的filters
3.在build\darknet\x64\data\下，创建fre.names文件，其中每一行写上一个你要训练的一个类别，eg：
apple
bickly
cell
bag
4.在build\darknet\x64\data\下，创建fre.data文件:
classes= 4
train = build/darknet/x64/data/train.txt # 这个下面说，train.txt里面存放着训练图片的存放路径
valid = build/darknet/x64/data/test.txt #
names = build/darknet/x64/data/fre.names # 步骤3创建的fre.names
backup = backup/
注：下面步骤5，6均用代码完成
5.为每一张训练图片创建一个.txt文件。 例如 img1.jpg，那么你就创建img1.txt，这个文件里面包含：
1 0.716797 0.395833 0.216406 0.147222
0 0.687109 0.379167 0.255469 0.158333
1 0.420312 0.395833 0.140625 0.166667
注意，这个创建的img1.txt与训练图片img1.jpg在同一个文件夹里，eg：build/darknet/x64/data/obj
class_id、x_center 、y_center、width、height
注意上面的x_center y_center width height的数值都是相对图片的尺度而言的，即：
x_center = x / image_width
y_center = y / image_height
width = gt_width / image_width
height = gt_height / image_height
那么问题来了，如何将labelimg的标签（xml文件）转换为yolov4的标签格式，我们通过代码来实现，但是先不急，先把步骤6讲完，再用代码一起实现步骤5、6。
6.在build/darknet/x64/data/下，创建train.txt文件，里面保存训练图片文件的路径：
data/obj/img1.jpg
data/obj/img2.jpg
data/obj/img3.jpg
好了，下面奉上代码：xml_to_txt.py
# -*- coding: utf-8 -*-
# ——author—— = “调得一手好参”
# Email : 928343544@qq.com
# time : 2020.5.12
# function: 将xml文件转为yolo的标签import os
import argparse
import xml.etree.ElementTree as ET
import globdef xml_to_txt(data_path,anno_path,path,use_difficult_bbox=False):classes = ['am','fm','gsm','qpsk']image_inds = file_name(data_path+"train_label/") #遍历xml文件with open(anno_path, 'a') as f:for image_ind in image_inds:img_txt = data_path + 'obj/'+ image_ind + '.txt'img_txt_file = open(img_txt, 'w')image_path = os.path.join(data_path, 'obj/', image_ind + '.jpg') label_path = os.path.join(data_path, 'train_label', image_ind +'.xml')root = ET.parse(label_path).getroot()objects_size = root.findall('size')image_width = int(objects_size[0].find('width').text)image_height = int(objects_size[0].find('height').text)objects = root.findall('object')for obj in objects:difficult = obj.find('difficult').text.strip()if (not use_difficult_bbox) and(int(difficult) == 1):continuebbox = obj.find('bndbox')class_ind = str(classes.index(obj.find('name').text.lower().strip()))xmin = int(bbox.find('xmin').text.strip())xmax = int(bbox.find('xmax').text.strip())ymin = int(bbox.find('ymin').text.strip())ymax = int(bbox.find('ymax').text.strip())x_center = str((xmin + xmax)/(2*image_width))y_center = str((ymin + ymax)/(2*image_height))width_ = str((xmax - xmin)/(image_width))height_ = str((ymax - ymin)/(image_height))class_ind += ' ' + ','.join([x_center+' '+y_center+' '+width_+' '+height_])img_txt_file.write(class_ind + "\n")f.write(image_path + "\n")def file_name(file_dir):L = []for root, dirs, files in os.walk(file_dir):for file in files:if os.path.splitext(file)[1] == '.xml':L.append(file.split(".")[0])return Lif __name__ == '__main__':num1 = xml_to_txt('./build/darknet/x64/data/','./build/darknet/x64/data/train.txt','train')# num2 = convert_voc_annotation('./data/', './data/test.txt', False)print('done')
快速使用代码：
1)将xml_to_txt.py放在darknet-master目录下
2)将xml文件放在 ./build/darknet/x64/data/train_label
3)将训练图片文件放在 ./build/darknet/x64/data/obj
4)运行代码
在./build/darknet/x64/data/下生成 train.txt文件
在./build/darknet/x64/data/obj/ 下生成 img.txt
上面是将xml文件转为txt文件，如果你使用labelme软件打.json文件标签，用下面的代码转为txt文件：
import os
import argparse
import json
import glob
import numpy as npdef json_to_txt(data_path,save_path,train_txt_file):classes = ['apple','balance','car','person'] #这里填写你要训练的类train_txt = train_txt_filetrain_txt_path = open(train_txt, 'w')for num,json_file in enumerate(data_path):with open(json_file,'r') as fp: data = json.load(fp) # 加载json文件file_name = ''file_name = ''for name in data["imagePath"].split('.'):if name == 'jpg' or name == 'png':breakelif name[-1]=='Z':file_name +=nameelif data["imagePath"].split('.').index(name)<2:file_name += name+'.'else:file_name += nameif file_name[-1] =='.':img_txt = save_path + file_name+ 'txt'image_path = save_path + file_name+ 'jpg'else:img_txt = save_path + file_name+ '.txt'image_path = save_path + file_name+ '.jpg'with open(img_txt,'w') as img_txt_file:img_height = data['imageHeight']img_width = data['imageWidth']for obj in data["shapes"]:class_id = str(classes.index(obj["label"]))x_zoom = []y_zoom = []for zoom in obj["points"]:x_zoom.append(zoom[0])y_zoom.append(zoom[1])min_x = min(x_zoom)max_x = max(x_zoom)min_y = min(y_zoom)max_y = max(y_zoom)# # convert()# right_low_index = np.argmax(np.sum(np.array(obj["points"]),1))# left_top_index = np.argmin(np.sum(np.array(obj["points"]),1))# right_low = obj["points"][right_low_index]# left_top = obj["points"][left_top_index]x_center = str((min_x + max_x)/(2*img_width))y_center = str((min_y + max_y)/(2*img_height))width_ = str(abs(max_x -min_x)/(img_width))height_ = str(abs(max_y - min_y)/(img_height))class_id += ' ' + ','.join([x_center+' '+y_center+' '+width_+' '+height_])img_txt_file.write(class_id + "\n")train_txt_path.write(image_path + "\n")if __name__ == '__main__':###训练数据生成labelme_train_json=glob.glob('存放训练json文件的路径/*.json')train_txt_file = 'train.txt' #文件路径save_train_path = '设置生成的训练标签存放的路径'json_to_txt(labelme_train_json,save_train_path,train_txt_file)###测试数据生成labelme_test_json=glob.glob('存放测试json文件的路径/*.json')test_txt_file = 'test.txt' #文件路径save_test_path = '设置生成的测试标签存放的路径'json_to_txt(labelme_test_json,save_test_path ,test_txt_file )
7. 训练：(linux)
./darknet detector train build/darknet/x64/data/fre.data cfg/yolo-obj_fre.cfg build/darknet/x64/yolov4.conv.137
windows下只需要将上面的./darknet改为darknet.exe
