VisionEnsembleModel.py

# Contenido de: VisionEnsembleModel.py

import torch
import torch.nn as nn
import timm

class VisionEnsembleModel(nn.Module):
    def __init__(self, num_classes, cnn_model_name='efficientnet_b2', vit_model_name='vit_small_patch16_224'):
        super().__init__()
        self.cnn = timm.create_model(cnn_model_name, pretrained=False, num_classes=num_classes)
        cnn_features = self.cnn.get_classifier().in_features
        self.cnn.reset_classifier(0)

        self.vit = timm.create_model(vit_model_name, pretrained=False, num_classes=num_classes)
        vit_features = self.vit.head.in_features
        self.vit.head = nn.Identity()

        self.classifier = nn.Sequential(
            nn.BatchNorm1d(cnn_features + vit_features),
            nn.Linear(cnn_features + vit_features, 512),
            nn.ReLU(),
            nn.Dropout(0.5),
            nn.Linear(512, num_classes)
        )

    def forward(self, image):
        cnn_feat = self.cnn(image)
        vit_feat = self.vit(image)
        combined = torch.cat([cnn_feat, vit_feat], dim=1)
        output = self.classifier(combined)
        return output