# 导入必要的库 import os # 文件和路径管理 import time # 计时功能 import logging # 日志记录 import json # 处理JSON数据 import numpy as np # 数值运算 from PIL import Image # 图像处理 import torch # PyTorch深度学习框架 import torch.nn as nn # 神经网络模块 import torch.optim as optim # 优化器 from torch.utils.data import Dataset, DataLoader # 数据集和加载器 import torchvision.utils as vutils # 可视化工具 # COCO数据集相关库 from pycocotools import mask as maskUtils # COCO掩码处理工具 import albumentations as A # 数据增强库 from albumentations.pytorch import ToTensorV2 # Albumentations到Tensor的转换 torch.backends.cudnn.benchmark = True # 固定随机种子，确保实验可复现 torch.manual_seed(42) # 配置日志格式 logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s') # ------------------ 定义 U-Net 模型（自定义） ------------------ class UNet(nn.Module): def __init__(self, in_channels=3, base_channels=64, out_channels=1): super(UNet, self).__init__() # 双卷积块定义 def double_conv(in_c, out_c): return nn.Sequential( nn.Conv2d(in_c, out_c, kernel_size=3, padding=1), nn.BatchNorm2d(out_c), nn.ReLU(inplace=True), nn.Conv2d(out_c, out_c, kernel_size=3, padding=1), nn.BatchNorm2d(out_c), nn.ReLU(inplace=True) ) # 下采样路径 self.enc1 = double_conv(in_channels, base_channels) self.enc2 = double_conv(base_channels, base_channels*2) self.enc3 = double_conv(base_channels*2, base_channels*4) self.enc4 = double_conv(base_channels*4, base_channels*8) self.pool = nn.MaxPool2d(2) # 最大池化层 # 中心层 self.center = double_conv(base_channels*8, base_channels*16) # 上采样路径 self.up4 = nn.ConvTranspose2d(base_channels*16, base_channels*8, kernel_size=2, stride=2) self.dec4 = double_conv(base_channels*16, base_channels*8) self.up3 = nn.ConvTranspose2d(base_channels*8, base_channels*4, kernel_size=2, stride=2) self.dec3 = double_conv(base_channels*8, base_channels*4) self.up2 = nn.ConvTranspose2d(base_channels*4, base_channels*2, kernel_size=2, stride=2) self.dec2 = double_conv(base_channels*4, base_channels*2) self.up1 = nn.ConvTranspose2d(base_channels*2, base_channels, kernel_size=2, stride=2) self.dec1 = double_conv(base_channels*2, base_channels) # 最终1x1卷积层 self.final = nn.Conv2d(base_channels, out_channels, kernel_size=1) def forward(self, x): # 编码路径 e1 = self.enc1(x) e2 = self.enc2(self.pool(e1)) e3 = self.enc3(self.pool(e2)) e4 = self.enc4(self.pool(e3)) # 中心层 c = self.center(self.pool(e4)) # 解码路径 d4 = self.up4(c) d4 = torch.cat([d4, e4], dim=1) # 特征拼接 d4 = self.dec4(d4) d3 = self.up3(d4) d3 = torch.cat([d3, e3], dim=1) d3 = self.dec3(d3) d2 = self.up2(d3) d2 = torch.cat([d2, e2], dim=1) d2 = self.dec2(d2) d1 = self.up1(d2) d1 = torch.cat([d1, e1], dim=1) d1 = self.dec1(d1) return self.final(d1) # 最终输出 # ------------------ Dice Loss 与 IoU 指标 ------------------ class DiceLoss(nn.Module): def __init__(self, eps=1e-6): super(DiceLoss, self).__init__() self.eps = eps # 平滑项 def forward(self, logits, targets): probs = torch.sigmoid(logits) # 转换为概率 num = 2 * (probs * targets).sum(dim=(2,3)) + self.eps # 分子 den = probs.sum(dim=(2,3)) + targets.sum(dim=(2,3)) + self.eps # 分母 return 1 - (num/den).mean() # 返回Dice损失 # 计算IoU评分 def iou_score(preds, targets, eps=1e-6): preds = (preds > 0.5).float() # 二值化预测结果 inter = (preds * targets).sum(dim=(2,3)) # 交集 union = preds.sum(dim=(2,3)) + targets.sum(dim=(2,3)) - inter # 并集 return ((inter+eps)/(union+eps)).mean().item() # 返回IoU评分 # ------------------ COCO 分割数据集 ------------------ class CocoSegDataset(Dataset): def __init__(self, root_dir, annotation_file, transforms=None, mask_transforms=None): self.root_dir = root_dir # 数据集根目录 self.transforms = transforms # 数据变换 self.mask_transforms = mask_transforms # 掩码变换 with open(annotation_file, 'r') as f: self.coco = json.load(f) # 加载COCO标注文件 self.annotations = {} for ann in self.coco['annotations']: self.annotations.setdefault(ann['image_id'], []).append(ann) # 建立图像ID到标注的映射 self.image_ids = list(self.annotations.keys()) # 所有图像ID列表 def __len__(self): return len(self.image_ids) # 返回数据集大小 def __getitem__(self, idx): img_id = self.image_ids[idx] # 获取图像ID info = next(x for x in self.coco['images'] if x['id']==img_id) # 获取图像信息 img = Image.open(os.path.join(self.root_dir, info['file_name'])).convert('RGB') # 读取图像 h, w = info['height'], info['width'] # 获取图像尺寸 mask = np.zeros((h,w), dtype=np.uint8) # 初始化掩码 for ann in self.annotations[img_id]: # 生成掩码 seg = ann['segmentation'] if isinstance(seg, list): rle = maskUtils.merge(maskUtils.frPyObjects(seg,h,w)) else: rle = seg mask += maskUtils.decode(rle) mask = (mask>0).astype(np.float32) # 二值化掩码 mask = Image.fromarray(mask) # 转换为PIL图像 # 应用数据增强 if self.transforms and self.mask_transforms: aug = self.transforms(image=np.array(img), mask=np.array(mask)) img_t = aug['image']; m_t = aug['mask'].unsqueeze(0) else: img_t = ToTensorV2()(image=np.array(img))['image'] m_t = ToTensorV2()(image=np.array(mask))['image'] return img_t, m_t # 返回处理后的图像和掩码 # ------------------ 主训练流程 ------------------ if __name__ == '__main__': # 路径配置 train_dir, val_dir = '../data/train', '../data/valid' # 训练集和验证集目录 train_ann, val_ann = os.path.join(train_dir,'_annotations.coco.json'), os.path.join(val_dir,'_annotations.coco.json') # 标注文件路径 # 数据增强配置 train_tf = A.Compose([ A.Resize(256,256), # 调整尺寸 A.HorizontalFlip(0.5), # 水平翻转 A.RandomBrightnessContrast(0.2), # 随机亮度对比度 A.ShiftScaleRotate(0.5), # 随机位移缩放旋转 A.Normalize((0.485,0.456,0.406),(0.229,0.224,0.225)), # 归一化 ToTensorV2()]) # 转换为Tensor val_tf = A.Compose([ A.Resize(256,256), # 调整尺寸 A.Normalize((0.485,0.456,0.406),(0.229,0.224,0.225)), # 归一化 ToTensorV2()]) # 转换为Tensor # 数据加载 train_ds = CocoSegDataset(train_dir,train_ann,train_tf,train_tf) # 训练数据集 val_ds = CocoSegDataset(val_dir, val_ann, val_tf, val_tf) # 验证数据集 train_ld = DataLoader(train_ds,batch_size=8,shuffle=True,num_workers=4) # 训练数据加载器 val_ld = DataLoader(val_ds, batch_size=8,shuffle=False,num_workers=4) # 验证数据加载器 logging.info(f"Train samples: {len(train_ds)}, Val samples: {len(val_ds)}") # 输出数据集信息 # 模型与训练配置 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # 设备选择 print(f"当前设备：{device}") # 输出当前设备信息 model = UNet().to(device) # 初始化模型并移动到指定设备 opt = optim.AdamW(model.parameters(),lr=1e-3,weight_decay=1e-4) # 优化器 sched = optim.lr_scheduler.ReduceLROnPlateau(opt,'min',patience=3,factor=0.5) # 学习率调度器 bce = nn.BCEWithLogitsLoss(); dice = DiceLoss() # 损失函数 epochs=20; vis_dir='output/val_visuals'; os.makedirs(vis_dir,exist_ok=True) # 训练参数和可视化目录 # 训练循环 for ep in range(1,epochs+1): model.train(); run_loss=0 # 设置为训练模式 for imgs,msks in train_ld: imgs,msks=imgs.to(device),msks.to(device) # 将数据移动到指定设备 opt.zero_grad(); out=model(imgs) # 前向传播 l=(bce(out,msks)+dice(out,msks)); l.backward(); opt.step() # 计算损失、反向传播 run_loss+=l.item()*imgs.size(0) # 累计损失 tr_loss=run_loss/len(train_ds) # 计算平均训练损失 # 验证阶段 model.eval(); v_loss=0; v_iou=0; v_dice=0 with torch.no_grad(): for imgs,msks in val_ld: imgs,msks=imgs.to(device),msks.to(device) # 将数据移动到指定设备 out=model(imgs) # 前向传播 v_loss+=(bce(out,msks)+dice(out,msks)).item()*imgs.size(0) # 累计验证损失 pr=torch.sigmoid(out) # 转换为概率 v_iou+=iou_score(pr,msks)*imgs.size(0) # 累计IoU评分 v_dice+=(1 - dice(out,msks)).item()*imgs.size(0) # 累计Dice评分 v_loss/=len(val_ds); v_iou/=len(val_ds); v_dice/=len(val_ds) # 计算平均验证指标 # 输出训练信息 logging.info(f"Epoch {ep}/{epochs} - Tr:{tr_loss:.4f} Val:{v_loss:.4f} IoU:{v_iou:.4f} Dice:{v_dice:.4f}") # 可视化 si,sm=next(iter(val_ld)); si=si.to(device) # 获取示例图像 with torch.no_grad(): sp=torch.sigmoid(model(si)) # 进行预测 grid=vutils.make_grid(torch.cat([si.cpu(),sm.repeat(1,3,1,1).cpu(),sp.repeat(1,3,1,1).cpu()],0),nrow=si.size(0)) # 创建网格 vpth=os.path.join(vis_dir,f'ep{ep}.png'); vutils.save_image(grid,vpth) # 保存可视化结果 logging.info(f"Saved visual: {vpth}") # 输出保存信息 sched.step(v_loss) # 更新学习率 # 保存模型权重 torch.save(model.state_dict(),'unet_coco_segmentation.pth') logging.info('训练完成，模型已保存')