OOD分类项目训练

一、项目地址

GitHub – LooKing9218/UIOS

二、label制作

将训练、验证、测试数据的分类信息转换入.csv文件中，运行如下脚本即可：

import osimport csv #要读取的训练、验证、测试文件的目录，该文件下保存着以各个类别命名的文件夹和对应的分类图片root_path=r'/media/*********************/train' #类别种类classes=['cls1','cls2']def get_Write_file_infos(path):# 文件信息列表file_infos_list=[]typeclothes=os.listdir(path)for ii in typeclothes:everyfile=os.path.join(path , ii)for root, dirnames, filenames in os.walk(everyfile):for filename in filenames:file_infos = {}dirname=root #根据自己的需求更改路径地址filename1 ='train/'+ii+'/'+ filename#.split('.jpg')[0]flag = filename1[-1]file_infos["ImageId"] = filename1 file_infos["Flag"] = classes.index(ii)#将数据追加字典到列表中file_infos_list.append(file_infos)return file_infos_list#写入csv文件def write_csv(file_infos_list):with open('train_label.csv','a+',newline='') as csv_file_train:csv_writer = csv.DictWriter(csv_file_train,fieldnames=['ImageId','Flag'])csv_writer.writeheader()for each in file_infos_list:print(each)csv_writer.writerow(each)def main():file_infos_list =get_Write_file_infos(root_path)write_csv(file_infos_list)if __name__ == '__main__':main()print('The End！')

生成情况如下：

三、运行程序

（1）修改参数文件 utils/config.py

# -*- coding: utf-8 -*-class DefaultConfig(object):net_work = 'ResUnNet50'num_classes = 2num_epochs = 100batch_size = 256validation_step = 1root = "/media/code/"train_file = "train_label.csv"val_file = "val_label.csv"test_file = "test_label.csv"lr = 1e-4lr_mode = 'poly'momentum = 0.9weight_decay = 1e-4save_model_path = './Model_Saved'.format(net_work,lr)log_dirs = './Logs_Adam_0304'pretrained =True# Falsepretrained_model_path ='/media/code/UIOS-master/Trained/archive/data/99843712' #Nonecuda = 0num_workers = 4use_gpu = Truetrained_model_path = ''predict_fold = 'predict_mask'

(2)运行

命令：

python train.py

（3）运行界面

四、踩坑记录

问题原因：ValueError: Only one class present in y_true. ROC AUC score is not defined in that case.

解决方法：

（1）网上看了很多：

方法1：添加try-except

try:epoch_train_auc = metrics.roc_auc_score(labels, outputs)writer.add_scalar('Train/train_auc', float(epoch_train_auc),epoch)print('loss for train : {}，{}'.format(loss_train_mean,round(epoch_train_auc,6)))except ValueError:pass

方法2：DataLoader的参数设置shuffle=True

 train_loader = DataLoader(DatasetCFP(root=args.root,mode='train',data_file=args.train_file,),batch_size=args.batch_size, shuffle=True, pin_memory=True)val_loader = DataLoader(DatasetCFP(root=args.root,mode='val',data_file=args.val_file,),batch_size=args.batch_size, shuffle=True, pin_memory=True)test_loader = DataLoader(DatasetCFP(root=args.root,mode='test',data_file=args.test_file,),batch_size=args.batch_size, shuffle=True, pin_memory=True)

方法3：增大batch_size

(2)我的方法:

其实是我马虎大意

修改好config.py中的num_classes参数就行了，

见谅（不好意思~(￣▽￣)~*）