src/models/efficientnet.py

import torch
import torch.nn as nn
from timm import create_model

class EfficientNetClassifier(nn.Module):
    def __init__(self, model_name='efficientnet_b0', num_classes=5, pretrained=True, dropout=0.2):
        super().__init__()
        self.backbone = create_model(model_name, pretrained=pretrained, num_classes=0, global_pool='avg')
        with torch.no_grad():
            dummy = torch.randn(1, 3, 224, 224)
            self.feature_dim = self.backbone(dummy).shape[1]
        # DR分级头
        self.classifier_grading = nn.Sequential(
            nn.Dropout(dropout),
            nn.Linear(self.feature_dim, 512),
            nn.ReLU(inplace=True),
            nn.Dropout(dropout * 0.5),
            nn.Linear(512, num_classes)
        )
        # 二分类头
        self.classifier_diabetic = nn.Sequential(
            nn.Dropout(dropout),
            nn.Linear(self.feature_dim, 128),
            nn.ReLU(inplace=True),
            nn.Dropout(dropout * 0.5),
            nn.Linear(128, 1)
        )
        self._initialize_weights()
    def _initialize_weights(self):
        for m in list(self.classifier_grading) + list(self.classifier_diabetic):
            if isinstance(m, nn.Linear):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
                nn.init.constant_(m.bias, 0)
    def forward(self, x):
        features = self.backbone(x)
        out_grading = self.classifier_grading(features)
        out_diabetic = self.classifier_diabetic(features).squeeze(-1)  # shape: (B,)
        # 返回dict，兼容旧用法
        return {'grading': out_grading, 'diabetic': out_diabetic}
    def get_features(self, x):
        return self.backbone(x)