理论上用于分类的图像一般都不需要用labelme来标注的，笔者是因为刚好手上有这么一组数据，所以就顺带处理了。labelme标注完的数据每张还包含了一个json文件，这个在分类任务中用不上。具体的mmclassification使用方法在我的另一篇文章里有，需要注意的是现在分类任务被集合在mmpretrain里了。用法也优点区别，不过也都是细微的修改，都还能用。
数据处理代码如下：
import os, random, shutil# 最开始train的文件夹路径
ori_path = r'F:\Data\doctor_research\Arthritis_mini'# label种类：
label_list = ['arthritis', 'normal']# 抽取比例 ******自己改*******
val_ratio, test_ratio = 0.1, 0.2 # 抽取比例 ******自己改*******# 处理好的数据集路径
result_path = os.path.join(ori_path, 'result')# 存放图片的文件地址
train_path = os.path.join(result_path, 'train')
test_path = os.path.join(result_path, 'test')
val_path = os.path.join(result_path, 'val')# meta文件夹地址
meta_path = os.path.join(result_path, 'meta')
# meta文件夹下的txt文件
train_txt = os.path.join(meta_path, 'train.txt')
test_txt = os.path.join(meta_path, 'test.txt')
val_txt = os.path.join(meta_path, 'val.txt')# 预留内容
traintxt = []
testtxt = []
valtxt = []# 如果没有就创建
if not os.path.exists(result_path):os.mkdir(result_path)if not os.path.exists(meta_path):os.makedirs(meta_path)if not os.path.exists(train_path):os.makedirs(train_path)if not os.path.exists(test_path):os.makedirs(test_path)if not os.path.exists(val_path):os.makedirs(val_path)# 在train test val 文件夹中创建对应的label文件夹：
for label in label_list:print(f'label:{label}')os.makedirs(os.path.join(train_path, label), exist_ok=True)# print(f'train_path:{train_path}')os.makedirs(os.path.join(test_path, label), exist_ok=True)# print(f'test_path:{test_path}')os.makedirs(os.path.join(val_path, label), exist_ok=True)# print(f'val_path:{val_path}')ori_pic_path = os.path.join(ori_path, label)print(f'ori_pic_path is : {ori_pic_path}')# 将该label下所有的图像文件名暂存在一个list里面：temp_path = []for pic in os.listdir(ori_pic_path):# 判断是否为图片if pic.endswith('.jpg') or pic.endswith('.jpeg') or pic.endswith('.png'):# 先全部计入到一个临时list里：print(f'pic is {pic}')temp_path.append(pic)# print(f'temp_path is {temp_path}')# 计算该类别下的所有图片数量以及对应的验证集、测试集图片数量print(f'{label} num is {len(temp_path)}')# 验证集数量val_number = int(len(temp_path) * val_ratio)# 测试集数量test_number = int(len(temp_path) * test_ratio)# 抽取val数据集val_sample = random.sample(temp_path, val_number)print(f'val_number is:{val_number}')print(f'val_sample is:{val_sample}')# 把抽取的val数据剔除# temp_path.remove(val_sample)temp_path = [item for item in temp_path if item not in val_sample]# 对应的val文件夹：goal_val_path = os.path.join(result_path, 'val', label)# 将文件移动到val文件夹for name1 in val_sample:goal_name1_path = os.path.join(goal_val_path, name1)shutil.copy(os.path.join(ori_pic_path, name1), goal_name1_path)# 同时将该文件地址记录到val.txt中val_content = goal_name1_path + " " + str(label_list.index(label))valtxt.append(val_content)# with open(val_txt, 'w') as f:# f.write(goal_name1_path + " " + str(label_list.index(label)) + "\n")# 抽取test数据集test_sample = random.sample(temp_path, test_number)print(f'test_number is:{test_number}')# 把抽取的test数据剔除temp_path = [item for item in temp_path if item not in test_sample]# 对应的test文件夹goal_test_path = os.path.join(result_path, 'test', label)# 将文件移动到test文件夹for name2 in test_sample:goal_name2_path = os.path.join(goal_test_path, name2)shutil.copy(os.path.join(ori_pic_path, name2), goal_name2_path)# 同时将该文件地址记录到test.txt中# 同时将该文件地址记录到test.txt中test_content = goal_name2_path + " " + str(label_list.index(label))testtxt.append(test_content)# with open(test_txt, 'w') as f:# f.write(goal_name2_path + " " + str(label_list.index(label)) + "\n")# 把剩下的数据移入train数据集goal_train_path = os.path.join(result_path, 'train', label)print(f'goal_train_path is : {goal_train_path}')for name3 in temp_path:goal_name3_path = os.path.join(goal_train_path, name3)shutil.copy(os.path.join(ori_pic_path, name3), goal_name3_path)# 同时将该文件地址记录到test.txt中train_content = goal_name3_path + " " + str(label_list.index(label))traintxt.append(train_content)with open(train_txt, 'w') as f:for item in traintxt:f.write(item + '\n')with open(test_txt, 'w') as f:for item in testtxt:f.write(item + '\n')with open(val_txt, 'w') as f:for item in valtxt:f.write(item + '\n')
