# 解压PaddleSeg套件!unzip -oq /home/aistudio/data/data114849/PaddleSeg-release-2.3.zip登录后复制In [?]

# 为方便后期操作将文件夹改名!mv PaddleSeg-release-2.3 PaddleSeg登录后复制In [?]

# 安装依赖，AIStudio实验环境已有相关依赖#!pip install -r PaddleSeg/requirements.txt登录后复制

解压数据集

!unzip -oq /home/aistudio/data/data77571/train_and_label.zip -d data/!unzip -oq /home/aistudio/data/data77571/img_test.zip -d data/登录后复制

生成训练列表文件

这里直接按比例分割，没有用shuffle

import osimg_train_dir = os.listdir('./data/img_train')train_val_ratio = 0.9 #划分训练集和验证集的比例train_img_num = int(len(img_train_dir)*train_val_ratio)val_img_num = len(img_train_dir) - train_img_numtrain_list_origin = img_train_dir[:train_img_num]val_list_origin = img_train_dir[train_img_num:]print(len(train_list_origin),len(val_list_origin)) #输出训练集和验证集数量#生成trainlistwith open('train_list.txt','w') as f: for i in range(len(train_list_origin)): train_img_name = 'img_train/' + train_list_origin[i] train_lab_name = 'lab_train/' + train_list_origin[i].split('.')[0] + '.png' f.write(train_img_name + ' ' + train_lab_name + 'n')#生成vallistwith open('val_list.txt','w') as f: for i in range(len(val_list_origin)): val_img_name = 'img_train/' + val_list_origin[i] val_lab_name = 'lab_train/' + val_list_origin[i].split('.')[0] + '.png' f.write(val_img_name + ' ' + val_lab_name + 'n')登录后复制

59986 6666登录后复制

模型选择和实践

Swin Transformer简介

Swin Transformer是ViT的一大进步，建议参考李沐大神读论文的方法啃原论文

Swin Transformer: Hierarchical Vision Transformer using Shifted Windows