YOLOv8实例分割训练自己的数据集保姆级教程

1.利用labelme进行数据标注

1.1Labelme 安装方法

首先安装 Anaconda，然后运行下列命令：

1. ##################
2. ## for Python 2 ##
3. ##################
4. conda create --name=labelme python=2.7
5. source activate labelme
6. # conda install -c conda-forge pyside2
7. conda install pyqt
8. pip install labelme
9. # 如果想安装最新版本，请使用下列命令安装：
10. # pip install git+https://github.com/wkentaro/labelme.git
12. ##################
13. ## for Python 3 ##
14. ##################
15. conda create --name=labelme python=3.6
16. source activate labelme
17. # conda install -c conda-forge pyside2
18. # conda install pyqt
19. pip install pyqt5 # pyqt5 can be installed via pip on python3
20. pip install labelme
21. 输入以下指令打开
22. labelme

1.2Labelme 使用教程

使用 labelme 进行场景分割标注的教程详见：labelme

2.转换划分数据集

    对数据集进行转换和划分。注意：在数据标注的时候将图片和json文件放在不同的文件夹里。如下图所示，另外新建两个文件夹txt 和split。

 2.1将json格式文件转换为txt格式

新建json2txt.py文件，修改文件路径为自己的路径

1. # -*- coding: utf-8 -*-
2. import json
3. import os
4. import argparse
5. from tqdm import tqdm
8. def convert_label_json(json_dir, save_dir, classes):
9. json_paths = os.listdir(json_dir)
10. classes = classes.split(',')
12. for json_path in tqdm(json_paths):
13. # for json_path in json_paths:
14. path = os.path.join(json_dir, json_path)
15. with open(path, 'r') as load_f:
16. json_dict = json.load(load_f)
17. h, w = json_dict['imageHeight'], json_dict['imageWidth']
19. # save txt path
20. txt_path = os.path.join(save_dir, json_path.replace('json', 'txt'))
21. txt_file = open(txt_path, 'w')
23. for shape_dict in json_dict['shapes']:
24. label = shape_dict['label']
25. label_index = classes.index(label)
26. points = shape_dict['points']
28. points_nor_list = []
30. for point in points:
31. points_nor_list.append(point[0] / w)
32. points_nor_list.append(point[1] / h)
34. points_nor_list = list(map(lambda x: str(x), points_nor_list))
35. points_nor_str = ' '.join(points_nor_list)
37. label_str = str(label_index) + ' ' + points_nor_str + '\n'
38. txt_file.writelines(label_str)
41. if __name__ == "__main__":
42. """
43. python json2txt_nomalize.py --json-dir my_datasets/color_rings/jsons --save-dir my_datasets/color_rings/txts --classes "cat,dogs"
44. """
45. parser = argparse.ArgumentParser(description='json convert to txt params')
46. parser.add_argument('--json-dir', type=str,default='D:/ultralytics-main/data/json', help='json path dir')
47. parser.add_argument('--save-dir', type=str,default='D:/ultralytics-main/data/txt' ,help='txt save dir')
48. parser.add_argument('--classes', type=str, default='ccc,ccc1',help='classes')
49. args = parser.parse_args()
50. json_dir = args.json_dir
51. save_dir = args.save_dir
52. classes = args.classes
53. convert_label_json(json_dir, save_dir, classes)

 2.2划分数据集

新建split.py,修改文件路径为自己的路径

1. # 将图片和标注数据按比例切分为 训练集和测试集
2. import shutil
3. import random
4. import os
5. import argparse
8. # 检查文件夹是否存在
9. def mkdir(path):
10. if not os.path.exists(path):
11. os.makedirs(path)
13. def main(image_dir, txt_dir, save_dir):
14. # 创建文件夹
15. mkdir(save_dir)
16. images_dir = os.path.join(save_dir, 'images')
17. labels_dir = os.path.join(save_dir, 'labels')
19. img_train_path = os.path.join(images_dir, 'train')
20. img_test_path = os.path.join(images_dir, 'test')
21. img_val_path = os.path.join(images_dir, 'val')
23. label_train_path = os.path.join(labels_dir, 'train')
24. label_test_path = os.path.join(labels_dir, 'test')
25. label_val_path = os.path.join(labels_dir, 'val')
27. mkdir(images_dir);
28. mkdir(labels_dir);
29. mkdir(img_train_path);
30. mkdir(img_test_path);
31. mkdir(img_val_path);
32. mkdir(label_train_path);
33. mkdir(label_test_path);
34. mkdir(label_val_path);
36. # 数据集划分比例，训练集75%，验证集15%，测试集15%，按需修改
37. train_percent = 0.8
38. val_percent = 0.1
39. test_percent = 0.1
41. total_txt = os.listdir(txt_dir)
42. num_txt = len(total_txt)
43. list_all_txt = range(num_txt) # 范围 range(0, num)
45. num_train = int(num_txt * train_percent)
46. num_val = int(num_txt * val_percent)
47. num_test = num_txt - num_train - num_val
49. train = random.sample(list_all_txt, num_train)
50. # 在全部数据集中取出train
51. val_test = [i for i in list_all_txt if not i in train]
52. # 再从val_test取出num_val个元素，val_test剩下的元素就是test
53. val = random.sample(val_test, num_val)
55. print("训练集数目：{}, 验证集数目：{},测试集数目：{}".format(len(train), len(val), len(val_test) - len(val)))
56. for i in list_all_txt:
57. name = total_txt[i][:-4]
59. srcImage = os.path.join(image_dir, name + '.jpg')
60. srcLabel = os.path.join(txt_dir, name + '.txt')
62. if i in train:
63. dst_train_Image = os.path.join(img_train_path, name + '.jpg')
64. dst_train_Label = os.path.join(label_train_path, name + '.txt')
65. shutil.copyfile(srcImage, dst_train_Image)
66. shutil.copyfile(srcLabel, dst_train_Label)
67. elif i in val:
68. dst_val_Image = os.path.join(img_val_path, name + '.jpg')
69. dst_val_Label = os.path.join(label_val_path, name + '.txt')
70. shutil.copyfile(srcImage, dst_val_Image)
71. shutil.copyfile(srcLabel, dst_val_Label)
72. else:
73. dst_test_Image = os.path.join(img_test_path, name + '.jpg')
74. dst_test_Label = os.path.join(label_test_path, name + '.txt')
75. shutil.copyfile(srcImage, dst_test_Image)
76. shutil.copyfile(srcLabel, dst_test_Label)
79. if __name__ == '__main__':
80. """
81. python split_datasets.py --image-dir my_datasets/color_rings/imgs --txt-dir my_datasets/color_rings/txts --save-dir my_datasets/color_rings/train_data
82. """
83. parser = argparse.ArgumentParser(description='split datasets to train,val,test params')
84. parser.add_argument('--image-dir', type=str,default='D:/ultralytics-main/data', help='image path dir')
85. parser.add_argument('--txt-dir', type=str,default='D:/ultralytics-main/data/txt' , help='txt path dir')
86. parser.add_argument('--save-dir', default='D:/ultralytics-main/data/split',type=str, help='save dir')
87. args = parser.parse_args()
88. image_dir = args.image_dir
89. txt_dir = args.txt_dir
90. save_dir = args.save_dir
92. main(image_dir, txt_dir, save_dir)

运行完后得到如下文件

3.训练设置

3.1新建seg.yaml文件 ，按照下列格式创建   我一般写成绝对路径，方便一点。

1. train: D:\ultralytics-main\data\split\images\train # train images (relative to 'path') 128 images
2. val: D:\ultralytics-main\data\split\images\val # val images (relative to 'path') 128 images
3. test: D:\ultralytics-main\data\split\images\test # test images (optional)
5. # Classes
6. names:
7. 0: ccc
8. 1: ccc1

3.2训练参数设置

1. task: segment # YOLO task, i.e. detect, segment, classify, pose
2. mode: train # YOLO mode, i.e. train, val, predict, export, track, benchmark
4. # Train settings -------------------------------------------------------------------------------------------------------
5. model: yolov8s-seg.yaml # path to model file, i.e. yolov8n.pt, yolov8n.yaml
6. #model:runs/detect/yolov8s/weights/best.pt
7. data: seg.yaml # path to data file, i.e. coco128.yaml
8. epochs: 10 # number of epochs to train for
9. patience: 50 # epochs to wait for no observable improvement for early stopping of training
10. batch: 16 # number of images per batch (-1 for AutoBatch)

然后开始训练即可。

参考：

(52条消息) 数据标注软件labelme详解_黑暗星球的博客-CSDN博客

(52条消息) YOLOv5-7.0实例分割训练自己的数据，切分mask图并摆正_yolo 图像分割_jin__9981的博客-CSDN博客