3. 对抗性域适应网络构建
任务说明
本案例构建一个对抗性域适应网络，用于跨域目标检测中的特征对齐。源域为合成图像（如Sim10K风格），目标域为真实图像（如Cityscapes风格），通过域判别器以对抗方式促使特征提取器学习域不变表示。最终在目标域上评估检测性能，并与无域适应基线对比。
网络架构与对抗机制
整体采用Faster R-CNN作为检测骨干，在特征提取层之后引入梯度反转层（GRL）和域判别器。前向传播时，域判别器接收来自源域和目标域的特征图，输出域类别概率；反向传播时，GRL将梯度取反，使特征提取器朝着混淆域判别器的方向优化，从而对齐跨域特征分布。检测损失与域对抗损失联合训练，总损失为：
L = L_det + λ · L_adv其中λ为对抗损失权重，本案例设为0.1。域判别器结构为三层卷积加全连接，输入特征图尺寸为14×14×1024。
核心代码实现
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models.detection import fasterrcnn_resnet50_fpn
class GradientReversalLayer(torch.autograd.Function):
"""梯度反转层：前向恒等，反向梯度取反"""
@staticmethod
def forward(ctx, x, lambda_val):
ctx.lambda_val = lambda_val
return x.clone()
@staticmethod
def backward(ctx, grad_output):
return -ctx.lambda_val * grad_output, None
def grad_reverse(x, lambda_val=0.1):
return GradientReversalLayer.apply(x, lambda_val)
class DomainDiscriminator(nn.Module):
"""域判别器：输入特征图，输出域类别概率"""
def __init__(self, in_channels=1024):
super().__init__()
self.conv1 = nn.Conv2d(in_channels, 512, kernel_size=3, padding=1)
self.conv2 = nn.Conv2d(512, 256, kernel_size=3, padding=1)
self.conv3 = nn.Conv2d(256, 128, kernel_size=3, padding=1)
self.pool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(128, 2)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = self.pool(x).flatten(1)
return self.fc(x)
class AdversarialDomainAdaptation(nn.Module):
"""对抗性域适应检测网络"""
def __init__(self, num_classes=8, lambda_adv=0.1):
super().__init__()
# 使用Faster R-CNN作为检测骨干
self.detector = fasterrcnn_resnet50_fpn(pretrained=True)
# 替换分类头为自定义类别数（含背景）
in_features = self.detector.roi_heads.box_predictor.cls_score.in_features
self.detector.roi_heads.box_predictor = \
torchvision.models.detection.faster_rcnn.FastRCNNPredictor(
in_features, num_classes)
# 域判别器
self.domain_disc = DomainDiscriminator(in_channels=1024)
self.lambda_adv = lambda_adv
def forward(self, images, targets=None, domain_labels=None):
"""
images: list[Tensor] 输入图像
targets: list[dict] 检测标注（源域提供，目标域可为None）
domain_labels: Tensor 域标签（0=源域, 1=目标域）
返回: 训练时(dict), 推理时检测结果
"""
# 提取特征（使用backbone的body部分）
features = self.detector.backbone(images)
# 取最高分辨率特征图（FPN的P2层）
feat = features['0'] # [B, 256, H, W]
# 域对抗分支
domain_loss = None
if domain_labels is not None:
# 梯度反转后送入域判别器
feat_reversed = grad_reverse(feat, self.lambda_adv)
domain_logits = self.domain_disc(feat_reversed)
domain_loss = F.cross_entropy(domain_logits, domain_labels)
# 检测分支
if self.training:
assert targets is not None, "训练时必须提供检测标注"
loss_dict = self.detector(images, targets)
if domain_loss is not None:
loss_dict['domain_loss'] = domain_loss
return loss_dict
else:
return self.detector(images)
训练流程与数据加载
def train_one_epoch(model, source_loader, target_loader, optimizer, device):
"""单轮训练：交替从源域和目标域采样"""
model.train()
total_loss = 0.0
for (src_images, src_targets), (tgt_images, _) in zip(source_loader, target_loader):
# 合并源域和目标域图像
images = list(src_images) + list(tgt_images)
images = [img.to(device) for img in images]
# 域标签：源域=0，目标域=1
batch_size = len(src_images)
domain_labels = torch.cat([
torch.zeros(batch_size, dtype=torch.long),
torch.ones(len(tgt_images), dtype=torch.long)
]).to(device)
# 源域有检测标注，目标域无标注
src_targets = [{k: v.to(device) for k, v in t.items()} for t in src_targets]
# 前向与损失
loss_dict = model(images, src_targets, domain_labels)
loss = sum(loss_dict.values())
# 反向传播
optimizer.zero_grad()
loss.backward()
optimizer.step()
total_loss += loss.item()
return total_loss / len(source_loader)
数据准备与运行
本案例使用简化版Sim10K与Cityscapes数据接口，实际运行时需替换为真实数据集路径。以下为最小可运行的数据加载示例：
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
class SimpleDomainDataset(Dataset):
"""简化域数据集：返回图像和标注"""
def __init__(self, image_paths, domain, has_annotations=True):
self.image_paths = image_paths
self.domain = domain # 'source' 或 'target'
self.has_annotations = has_annotations
self.transform = transforms.Compose([
transforms.Resize((800, 800)),
transforms.ToTensor(),
])
def __len__(self):
return len(self.image_paths)
def __getitem__(self, idx):
from PIL import Image
img = Image.open(self.image_paths[idx]).convert('RGB')
img = self.transform(img)
if self.has_annotations:
# 简化标注：假设每张图有一个目标框
target = {
'boxes': torch.tensor([[100, 100, 300, 300]], dtype=torch.float32),
'labels': torch.tensor([1], dtype=torch.int64),
}
return img, target
else:
return img, None
# 使用示例（需替换为实际路径）
# source_paths = ['sim10k_1.jpg', 'sim10k_2.jpg', ...]
# target_paths = ['cityscapes_1.jpg', 'cityscapes_2.jpg', ...]
# source_loader = DataLoader(SimpleDomainDataset(source_paths, 'source'), batch_size=2)
# target_loader = DataLoader(SimpleDomainDataset(target_paths, 'target', has_annotations=False), batch_size=2)
# 初始化与训练
# device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# model = AdversarialDomainAdaptation(num_classes=8).to(device)
# optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
# for epoch in range(10):
# loss = train_one_epoch(model, source_loader, target_loader, optimizer, device)
# print(f'Epoch {epoch}: loss = {loss:.4f}')
评估方法与判定标准
在目标域测试集上评估模型性能，使用mAP（mean Average Precision）作为指标。评估流程：
def evaluate(model, test_loader, device):
"""在目标域测试集上评估mAP"""
model.eval()
all_preds = []
all_targets = []
with torch.no_grad():
for images, targets in test_loader:
images = [img.to(device) for img in images]
predictions = model(images) # 推理模式
for pred, tgt in zip(predictions, targets):
# 提取预测框和标签
pred_boxes = pred['boxes'].cpu()
pred_scores = pred['scores'].cpu()
pred_labels = pred['labels'].cpu()
# 保留高置信度预测
keep = pred_scores > 0.5
all_preds.append({
'boxes': pred_boxes[keep],
'scores': pred_scores[keep],
'labels': pred_labels[keep],
})
all_targets.append(tgt)
# 计算mAP（需使用pycocotools或自定义实现）
# from pycocotools.coco import COCO
# from pycocotools.cocoeval import COCOeval
# ... 标准COCO评估流程
# 返回mAP值
return mAP
# 判定标准：
# 1. 对抗训练后，目标域mAP应高于无域适应基线（仅源域训练）
# 2. 域判别器准确率应趋近于50%（表示特征已无法区分域）
# 3. 训练收敛时，总损失应稳定下降
验证路径
完成训练的验证步骤：
- 运行训练脚本，观察损失曲线：检测损失应下降，域对抗损失应稳定在0.7附近（接近随机猜测的交叉熵）
- 在目标域测试集上运行
evaluate()，记录mAP值 - 对比实验：仅用源域数据训练相同骨干网络（无对抗分支），在目标域上评估得到基线mAP
- 若对抗训练后的mAP高于基线，且域判别器准确率接近50%，则判定域适应有效
注意：本案例代码为教学演示的最小实现，实际应用需替换为完整数据集加载、数据增强和更充分的训练轮次。上述代码中的标注格式与Faster R-CNN要求一致，可直接运行训练循环。