src/models/vit.py

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

class VisionTransformerClassifier(nn.Module):
    def __init__(self, model_name='vit_small_patch16_224', num_classes=5, pretrained=True, dropout=0.2):
        super().__init__()
        self.backbone = create_model(model_name, pretrained=pretrained, num_classes=0)
        with torch.no_grad():
            dummy = torch.randn(1, 3, 224, 224)
            self.feature_dim = self.backbone(dummy).shape[1]
        self.classifier = nn.Sequential(
            nn.LayerNorm(self.feature_dim),
            nn.Dropout(dropout),
            nn.Linear(self.feature_dim, 256),
            nn.GELU(),
            nn.Dropout(dropout * 0.5),
            nn.Linear(256, num_classes)
        )
        self._initialize_weights()
    def _initialize_weights(self):
        for m in self.classifier:
            if isinstance(m, nn.Linear):
                nn.init.trunc_normal_(m.weight, std=0.02)
                nn.init.constant_(m.bias, 0)
    def forward(self, x):
        features = self.backbone(x)
        return self.classifier(features)
    def get_features(self, x):
        return self.backbone(x)