| import torch.nn as nn | |
| class LesionEncoder(nn.Module): | |
| def __init__(self, input_channels=1, feature_dim=128): | |
| super().__init__() | |
| self.encoder = nn.Sequential( | |
| nn.Conv2d(input_channels, 32, kernel_size=3, padding=1), | |
| nn.ReLU(inplace=True), | |
| nn.Conv2d(32, 64, kernel_size=3, padding=1), | |
| nn.ReLU(inplace=True), | |
| nn.AdaptiveAvgPool2d((1, 1)), # 输出 [B, 64, 1, 1] | |
| nn.Flatten(), # [B, 64] | |
| nn.Linear(64, feature_dim), # → [B, 128] | |
| nn.ReLU(inplace=True) | |
| ) | |
| def forward(self, x): # x: [B, 1, H, W] | |
| return self.encoder(x) # [B, 128] | |