【飞桨学习赛：遥感影像地块分割】22年7月第一名方案

# !git clone https://github.com/lzzzzzm/AI-Studio-LrCompetition-Segmentation-of-sensing-imageblock!unzip -q data/data162694/PaddleSeg-SenseingBlock.zip -d ./

!unzip -q data/data80164/train_and_label.zip -d ~/data/
!unzip -q data/data80164/img_test.zip -d ~/data/

import osimport numpy as np

datas = []
image_base = 'data/img_train'   # 训练集原图路径annos_base = 'data/lab_train'   # 训练集标签路径# 读取原图文件名ids_ = [v.split('.')[0] for v in os.listdir(image_base)]# 将训练集的图像集和标签路径写入datas中for id_ in ids_:
    img_pt0 = os.path.join(image_base, '{}.jpg'.format(id_))
    img_pt1 = os.path.join(annos_base, '{}.png'.format(id_))
    datas.append((img_pt0.replace('/home/aistudio', ''), img_pt1.replace('/home/aistudio', '')))    if os.path.exists(img_pt0) and os.path.exists(img_pt1):        pass
    else:        raise "path invalid!"# 打印datas的长度和具体存储例子print('total:', len(datas))print(datas[0][0])print(datas[0][1])print(datas[10][:])

import numpy as np# 四类标签，这里用处不大，比赛评测是以0、1、2、3类来对比评测的labels = ['建筑', '耕地', '林地',  '其他']# 将labels写入标签文件with open('labels.txt', 'w') as f:    for v in labels:
        f.write(v+'\n')# 随机打乱datasnp.random.seed(5)
np.random.shuffle(datas)# 验证集与训练集的划分，0.05表示5%为训练集，95%为训练集split_num = int(0.05*len(datas))# 划分训练集和验证集train_data = datas[:-split_num]
valid_data = datas[-split_num:]# 写入训练集listwith open('train_list.txt', 'w') as f:    for img, lbl in train_data:
        f.write(img + ' ' + lbl + '\n')# 写入验证集listwith open('valid_list.txt', 'w') as f:    for img, lbl in valid_data:
        f.write(img + ' ' + lbl + '\n')# 打印训练集和测试集大小print('train:', len(train_data))print('valid:', len(valid_data))

import cv2 as cvimport tqdm# 写入包含类别3的采样样本sample_list = []with open('train_list.txt', 'r') as f:
    lines = f.readlines()    for line in tqdm.tqdm(lines):
        image_path, label_path = line.split(' ')
        label_path = label_path[:-1]
        label = cv.imread(label_path, cv.IMREAD_GRAYSCALE)        if label is not None:            if (label==3).any() == True:
                sample_list.append(line)
f.close()with open('sample.txt', 'w') as f:    for line in sample_list:
        f.writelines(line)
f.close()

!python AI-Studio-LrCompetition-Segmentation-of-sensing-imageblock/train.py \
    --config AI-Studio-LrCompetition-Segmentation-of-sensing-imageblock/myconfig/segformer-b2-rs.yml \
    --use_vdl \
    --do_eval \
    --num_workers 4 \
    --save_interval 10000 \
    --save_dir segformer-b2-rs \
    --log_iters 2000

!python AI-Studio-LrCompetition-Segmentation-of-sensing-imageblock/predict.py \
       --config AI-Studio-LrCompetition-Segmentation-of-sensing-imageblock/myconfig/segformer-b2-rs.yml \
       --model_path segformer-b2-rs/model.pdparams \
       --aug_pred \
       --scales 0.75 1.0 1.25 \
       --image_path data/img_testA \
       --custom_color 0 0 0 1 1 1 2 2 2 3 3 3 \
       --save_dir result/SegFormer-B2/

import osimport tqdmimport cv2

img_path = "/home/aistudio/result/SegFormer-B2/pseudo_color_prediction"rst_path =  "/home/aistudio/result/SegFormer-B2/result"if not os.path.exists(rst_path):
    os.makedirs(rst_path)
file_list = os.listdir(img_path)
img_list = [os.path.join(img_path, name) for name in file_list]for i,img_file in tqdm.tqdm(enumerate(img_list)):
    img = cv2.imread(img_file,cv2.IMREAD_GRAYSCALE)
    cv2.imwrite(os.path.join(rst_path,file_list[i]),img)

# 由预测结果生成提交文件!zip -r result.zip result/SegFormer-B2/result