标注工具 LabelMe 生成的标注文件为JSON格式，而YOLOv8中支持的为TXT文件格式。以下Python代码实现3个功能：
1.将JSON格式转换成TXT格式；
2.将数据集进行随机拆分，生成YOLOv8支持的目录结构；
3.生成YOLOv8支持的YAML文件。
代码test_labelme2yolov8.py如下：
import os
import json
import argparse
import colorama
import random
import shutildef parse_args():parser = argparse.ArgumentParser(description="json(LabelMe) to txt(YOLOv8)")parser.add_argument("--dir", required=True, type=str, help="images, json files, and generated txt files, all in the same directory")parser.add_argument("--labels", required=True, type=str, help="txt file that hold indexes and labels, one label per line, for example: face 0")parser.add_argument("--val_size", default=0.2, type=float, help="the proportion of the validation set to the overall dataset:[0., 0.5]")parser.add_argument("--name", required=True, type=str, help="the name of the dataset")args = parser.parse_args()return argsdef get_labels_index(name):labels = {} # key,valuewith open(name, "r") as file:for line in file:# print("line:", line)key_value = []for v in line.split(" "):# print("v:", v)key_value.append(v.replace("\n", "")) # remove line breaks(\n) at the end of the lineif len(key_value) != 2:print(colorama.Fore.RED + "Error: each line should have only two values(key value):", len(key_value))continuelabels[key_value[0]] = key_value[1]with open(name, "r") as file:line_num = len(file.readlines())if line_num != len(labels):print(colorama.Fore.RED + "Error: there may be duplicate lables:", line_num, len(labels))return labelsdef get_json_files(dir):jsons = []for x in os.listdir(dir):if x.endswith(".json"):jsons.append(x)return jsonsdef parse_json(name):with open(name, "r") as file:data = json.load(file)width = data["imageWidth"]height = data["imageHeight"]# print(f"width: {width}; height: {height}")objects=[]for shape in data["shapes"]:if shape["shape_type"] != "rectangle":print(colorama.Fore.YELLOW + "Warning: only the rectangle type is supported:", shape["shape_type"])continueobject = []object.append(shape["label"])object.append(shape["points"])objects.append(object)return width, height, objectsdef get_box_width_height(box):dist = lambda val: max(val) - min(val)x = [pt[0] for pt in box]y = [pt[1] for pt in box]return min(x), min(y), dist(x), dist(y)def bounding_box_normalization(width, height, objects, labels):boxes = []for object in objects:box = [] # class x_center y_center width heightbox.append(labels[object[0]])# print("point:", object[1])x_min, y_min, box_w, box_h = get_box_width_height(object[1])box.append(round((float(x_min + box_w / 2.0) / width), 6))box.append(round((float(y_min + box_h / 2.0) / height), 6))box.append(round(float(box_w / width), 6))box.append(round(float(box_h / height), 6))boxes.append(box)return boxes def write_to_txt(dir, json, width, height, objects, labels):boxes = bounding_box_normalization(width, height, objects, labels)# print("boxes:", boxes)name = json[:-len(".json")] + ".txt"# print("name:", name)with open(dir + "/" + name, "w") as file:for item in boxes:# print("item:", item)if len(item) != 5:print(colorama.Fore.RED + "Error: the length must be 5:", len(item))continuestring = item[0] + " " + str(item[1]) + " " + str(item[2]) + " " + str(item[3]) + " " + str(item[4]) + "\r"file.write(string)def json_to_txt(dir, jsons, labels):for json in jsons:name = dir + "/" + json# print("name:", name)width, height, objects = parse_json(name)# print(f"width: {width}; height: {height}; objects: {objects}")write_to_txt(dir, json, width, height, objects, labels)def is_in_range(value, a, b):return a <= value <= bdef get_random_sequence(length, val_size):numbers = list(range(0, length))val_sequence = random.sample(numbers, int(length*val_size))# print("val_sequence:", val_sequence)train_sequence = [x for x in numbers if x not in val_sequence]# print("train_sequence:", train_sequence)return train_sequence, val_sequencedef get_files_number(dir):count = 0for file in os.listdir(dir):if os.path.isfile(os.path.join(dir, file)):count += 1return countdef split_train_val(dir, jsons, name, val_size):if is_in_range(val_size, 0., 0.5) is False:print(colorama.Fore.RED + "Error: the interval for val_size should be:[0., 0.5]:", val_size)raisedst_dir_images_train = "datasets/" + name + "/images/train"dst_dir_images_val = "datasets/" + name + "/images/val"dst_dir_labels_train = "datasets/" + name + "/labels/train"dst_dir_labels_val = "datasets/" + name + "/labels/val"try:os.makedirs(dst_dir_images_train) #, exist_ok=Trueos.makedirs(dst_dir_images_val)os.makedirs(dst_dir_labels_train)os.makedirs(dst_dir_labels_val)except OSError as e:print(colorama.Fore.RED + "Error: cannot create directory:", e.strerror)raise# supported image formatsimg_formats = (".bmp", ".jpeg", ".jpg", ".png", ".webp")# print("jsons:", jsons)train_sequence, val_sequence = get_random_sequence(len(jsons), val_size)for index in train_sequence:for format in img_formats:file = dir + "/" + jsons[index][:-len(".json")] + format# print("file:", file)if os.path.isfile(file):shutil.copy(file, dst_dir_images_train)breakfile = dir + "/" + jsons[index][:-len(".json")] + ".txt"if os.path.isfile(file):shutil.copy(file, dst_dir_labels_train)for index in val_sequence:for format in img_formats:file = dir + "/" + jsons[index][:-len(".json")] + formatif os.path.isfile(file):shutil.copy(file, dst_dir_images_val)breakfile = dir + "/" + jsons[index][:-len(".json")] + ".txt"if os.path.isfile(file):shutil.copy(file, dst_dir_labels_val)num_images_train = get_files_number(dst_dir_images_train)num_images_val = get_files_number(dst_dir_images_val)num_labels_train = get_files_number(dst_dir_labels_train)num_labels_val = get_files_number(dst_dir_labels_val)if num_images_train + num_images_val != len(jsons) or num_labels_train + num_labels_val != len(jsons):print(colorama.Fore.RED + "Error: the number of files is inconsistent:", num_images_train, num_images_val, num_labels_train, num_labels_val, len(jsons))raisedef generate_yaml_file(labels, name):path = os.path.join("datasets", name, name+".yaml")# print("path:", path)with open(path, "w") as file:file.write("path: ../datasets/%s # dataset root dir\n" % name)file.write("train: images/train # train images (relative to 'path')\n")file.write("val: images/val # val images (relative to 'path')\n")file.write("test: # test images (optional)\n\n")file.write("# Classes\n")file.write("names:\n")for key, value in labels.items():# print(f"key: {key}; value: {value}")file.write(" %d: %s\n" % (int(value), key))if __name__ == "__main__":colorama.init()args = parse_args()# 1. parse JSON file and write it to a TXT filelabels = get_labels_index(args.labels)# print("labels:", labels)jsons = get_json_files(args.dir)# print("jsons:", jsons)json_to_txt(args.dir, jsons, labels)# 2. split the datasetsplit_train_val(args.dir, jsons, args.name, args.val_size)# 3. generate a YAML filegenerate_yaml_file(labels, args.name)print(colorama.Fore.GREEN + "====== execution completed ======")代码有些多，主要函数说明如下：
1.函数parse_args：解析输入参数；
2.函数get_labels_index：解析labels文件，数据集中的所有类别及对应的索引，格式labels.txt如下所示：生成YOLOv8的YAML文件时也需要此文件
face 0
hand 1
eye 2
mouth 3
horse 4
tree 5
bridge 6
house 73.函数get_json_files：获取指定目录下的所有json文件；
4.函数parse_json：解析json文件，将txt文件中需要的数据提取出来；
5.函数bounding_box_normalization：将bounding box值归一化到(0,1)区间；
6.函数write_to_txt：将最终结果写入txt文件；
7.函数split_train_val：将数据集随机拆分为训练集和验证集，并按YOLOv8支持的目录结构存放，根目录为datasets,接着是指定的数据集名，例如为fake，与YOLOv8中数据集coco8目录结构完全一致
8.函数generate_yaml_file：生成YOLOv8支持的yaml文件，存放在datasets/数据集名下，例如为fake.yaml
接收4个参数：参数dir为存放数据集的目录；参数labels指定labels文件；参数val_size指定验证集所占的比例；参数name指定新生成的YOLOv8数据集的名字
这里从网上随机下载了10幅图像，使用LabelMe进行了标注，执行结果如下图所示：
生成的fake.yaml文件如下图所示：
path: ../datasets/fake # dataset root dir
train: images/train # train images (relative to 'path')
val: images/val # val images (relative to 'path')
test: # test images (optional)# Classes
names:0: face1: hand2: eye3: mouth4: horse5: tree6: bridge7: house将生成的fake数据集进行训练，测试代码test_yolov8_detect.py如下：
import argparse
import colorama
from ultralytics import YOLOdef parse_args():parser = argparse.ArgumentParser(description="YOLOv8 object detect")parser.add_argument("--yaml", required=True, type=str, help="yaml file")parser.add_argument("--epochs", required=True, type=int, help="number of training")args = parser.parse_args()return argsdef train(yaml, epochs):model = YOLO("yolov8n.pt") # load a pretrained modelresults = model.train(data=yaml, epochs=epochs, imgsz=640) # train the modelmetrics = model.val() # It'll automatically evaluate the data you trained, no arguments needed, dataset and settings rememberedmodel.export(format="onnx", dynamic=True) # export the modelif __name__ == "__main__":colorama.init()args = parse_args()train(args.yaml, args.epochs)print(colorama.Fore.GREEN + "====== execution completed ======")执行结果如下图所示：目前此测试代码接收2个参数：参数yaml指定yaml文件；参数epochs指定训练次数；由以下结果可知，生成的新数据集无需做任何改动即可进行训练
GitHub：https://github.com/fengbingchun/NN_Test
