语义分割Baseline的基本流程

赛题背景

赛题链接
遥感技术已成为获取地表覆盖信息最为行之有效的手段，遥感技术已经成功应用于地表覆盖检测、植被面积检测和建筑物检测任务。本赛题使用航拍数据，需要参赛选手完成地表建筑物识别，将地表航拍图像素划分为有建筑物和无建筑物两类。

如下图，左边为原始航拍图，右边为对应的建筑物标注。

引入库

import numpy as np
import pandas as pd
import pathlib, sys, os, random, time
import cv2, gc
from tqdm import tqdm_notebook

import matplotlib.pyplot as plt
%matplotlib inline

import warnings
warnings.filterwarnings('ignore')

from tqdm.notebook import tqdm
import albumentations as A

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data as D
import torchvision
from torchvision import transforms as T

数据分析

赛题数据为航拍图，需要识别图片中的地表建筑具体像素位置。

• train_mask.csv：存储图片的标注的rle编码；
• train和test文件夹：存储训练集和测试集图片；

rle编码的具体的读取代码如下：

# 将图片编码为rle格式
def rle_encode(im):
    '''
    im: numpy array, 1 - mask, 0 - background
    Returns run length as string formated
    '''
    pixels = im.flatten(order = 'F')
    pixels = np.concatenate([[0], pixels, [0]])
    runs = np.where(pixels[1:] != pixels[:-1])[0] + 1
    runs[1::2] -= runs[::2]
    return ' '.join(str(x) for x in runs)

# 将rle格式进行解码为图片
def rle_decode(mask_rle, shape=(512, 512)):
    '''
    mask_rle: run-length as string formated (start length)
    shape: (height,width) of array to return 
    Returns numpy array, 1 - mask, 0 - background

    '''
    s = mask_rle.split()
    starts, lengths = [np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])]
    starts -= 1
    ends = starts + lengths
    img = np.zeros(shape[0]*shape[1], dtype=np.uint8)
    for lo, hi in zip(starts, ends):
        img[lo:hi] = 1
    return img.reshape(shape, order='F')

设置常用变量

• DEVICE：这是用于后续选择将数据放到GPU设备或者CPU设备上运行的属性
• IMAGE_SIZE：不同的图像大小，网络中的参数数量不一样。图像越大，参数越多，对算力要求越高。
• BATCH_SIZE: 批处理次数
• EPOCHES: 迭代轮数

DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu' 
EPOCHES = 20
BATCH_SIZE = 32
IMAGE_SIZE = 256

设置数据增强方式

这里用到了缩放、水平翻转、垂直翻转、随机90度旋转四类

trfm = A.Compose([
    A.Resize(IMAGE_SIZE, IMAGE_SIZE),
    A.HorizontalFlip(p=0.5),
    A.VerticalFlip(p=0.5),
    A.RandomRotate90(),
])

自定义数据集

class TianChiDataset(D.Dataset):
    def __init__(self, paths, rles, transform, test_mode=False):
        self.paths = paths
        self.rles = rles
        self.transform = transform
        self.test_mode = test_mode
        
        self.len = len(paths)
        self.as_tensor = T.Compose([
            T.ToPILImage(),
            T.Resize(IMAGE_SIZE),
            T.ToTensor(),
            T.Normalize([0.625, 0.448, 0.688],
                        [0.131, 0.177, 0.101]),
        ])
        
    # get data operation
    def __getitem__(self, index):
        img = cv2.imread(self.paths[index])
        if not self.test_mode:
            mask = rle_decode(self.rles[index])
            augments = self.transform(image=img, mask=mask)
            return self.as_tensor(augments['image']), augments['mask'][None]
        else:
            return self.as_tensor(img), ''        
    
    def __len__(self):
        """
        Total number of samples in the dataset
        """
        return self.len

加载训练数据

train_mask = pd.read_csv('data/train_mask.csv', sep='\t', names=['name', 'mask'])
train_mask['name'] = train_mask['name'].apply(lambda x: 'data/train/' + x)

dataset = TianChiDataset(
    train_mask['name'].values,
    train_mask['mask'].fillna('').values,
    trfm, False
)

把训练数据分为训练集和验证集

valid_idx, train_idx = [], []
for i in range(len(dataset)):
    if i % 7 == 0:
        valid_idx.append(i)
#     else:
    elif i % 7 == 1:
        train_idx.append(i)
        
train_ds = D.Subset(dataset, train_idx)
valid_ds = D.Subset(dataset, valid_idx)

# define training and validation data loaders
loader = D.DataLoader(
    train_ds, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)

vloader = D.DataLoader(
    valid_ds, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)

定义模型

使用的是fcn，特征提取是使用resnet50

def get_model():
    model = torchvision.models.segmentation.fcn_resnet50(True)
    
#     pth = torch.load("../input/pretrain-coco-weights-pytorch/fcn_resnet50_coco-1167a1af.pth")
#     for key in ["aux_classifier.0.weight", "aux_classifier.1.weight", "aux_classifier.1.bias", "aux_classifier.1.running_mean", "aux_classifier.1.running_var", "aux_classifier.1.num_batches_tracked", "aux_classifier.4.weight", "aux_classifier.4.bias"]:
#         del pth[key]
    
    model.classifier[4] = nn.Conv2d(512, 1, kernel_size=(1, 1), stride=(1, 1))
    return model

定义验证函数

@torch.no_grad()
def validation(model, loader, loss_fn):
    losses = []
    model.eval()
    for image, target in loader:
        image, target = image.to(DEVICE), target.float().to(DEVICE)
        output = model(image)['out']
        loss = loss_fn(output, target)
        losses.append(loss.item())
        
    return np.array(losses).mean()

定义损失函数

class SoftDiceLoss(nn.Module):
    def __init__(self, smooth=1., dims=(-2,-1)):

        super(SoftDiceLoss, self).__init__()
        self.smooth = smooth
        self.dims = dims
    
    def forward(self, x, y):
        tp = (x * y).sum(self.dims)
        fp = (x * (1 - y)).sum(self.dims)
        fn = ((1 - x) * y).sum(self.dims)
        
        dc = (2 * tp + self.smooth) / (2 * tp + fp + fn + self.smooth)
        dc = dc.mean()
        return 1 - dc
    
bce_fn = nn.BCEWithLogitsLoss()
dice_fn = SoftDiceLoss()

def loss_fn(y_pred, y_true):
    bce = bce_fn(y_pred, y_true)
    dice = dice_fn(y_pred.sigmoid(), y_true)
    return 0.8*bce+ 0.2*dice

加载模型，定义优化器，开始训练

model = get_model()
model.to(DEVICE);

optimizer = torch.optim.AdamW(model.parameters(),
                  lr=1e-4, weight_decay=1e-3)


header = r'''
        Train | Valid
Epoch |  Loss |  Loss | Time, m
'''
#          Epoch         metrics            time
raw_line = '{:6d}' + '\u2502{:7.3f}'*2 + '\u2502{:6.2f}'
print(header)

best_loss = 10
for epoch in range(1, EPOCHES+1):
    losses = []
    start_time = time.time()
    model.train()
    for image, target in tqdm_notebook(loader):
        
        image, target = image.to(DEVICE), target.float().to(DEVICE)
        optimizer.zero_grad()
        output = model(image)['out']
        loss = loss_fn(output, target)
        loss.backward()
        optimizer.step()
        losses.append(loss.item())
        # print(loss.item())
        
    vloss = validation(model, vloader, loss_fn)
    print(raw_line.format(epoch, np.array(losses).mean(), vloss,
                              (time.time()-start_time)/60**1))
    losses = []
    
    if vloss < best_loss:
        best_loss = vloss
        torch.save(model.state_dict(), 'model_best.pth')

加载最优模型，并在测试集上执行前向推理

trfm = T.Compose([
    T.ToPILImage(),
    T.Resize(IMAGE_SIZE),
    T.ToTensor(),
    T.Normalize([0.625, 0.448, 0.688],
                [0.131, 0.177, 0.101]),
])

subm = []

model.load_state_dict(torch.load("./model_best.pth"))
model.eval()

test_mask = pd.read_csv('data/test_a_samplesubmit.csv', sep='\t', names=['name', 'mask'])
test_mask['name'] = test_mask['name'].apply(lambda x: 'data/test_a/' + x)

for idx, name in enumerate(tqdm_notebook(test_mask['name'].iloc[:])):
    image = cv2.imread(name)
    image = trfm(image)
    with torch.no_grad():
        image = image.to(DEVICE)[None]
        score = model(image)['out'][0][0]
        score_sigmoid = score.sigmoid().cpu().numpy()
        score_sigmoid = (score_sigmoid > 0.5).astype(np.uint8)
        score_sigmoid = cv2.resize(score_sigmoid, (512, 512))

        
        # break
    subm.append([name.split('/')[-1], rle_encode(score_sigmoid)])

  0%|          | 0/2500 [00:00<?, ?it/s]

将预测结果保存到本地

subm = pd.DataFrame(subm)
subm.to_csv('./result.csv', index=None, header=None, sep='\t')

提交成绩