model.py

import torch
import torch.nn as nn
from torchvision.models import swin_t
from transformers import PretrainedConfig, PreTrainedModel

# 1. Define a Config class
class SwinClassifierConfig(PretrainedConfig):
    model_type = "swin_classifier"
    def __init__(self, num_classes=18, **kwargs):
        super().__init__(**kwargs)
        self.num_classes = num_classes

# 2. Update your Model class to inherit from PreTrainedModel
class SwinClassifier(PreTrainedModel):
    config_class = SwinClassifierConfig
    
    def __init__(self, config):
        super().__init__(config)
        # Use config.num_classes instead of a raw number
        self.backbone = swin_t()
        num_features = self.backbone.head.in_features
        
        self.backbone.head = nn.Sequential(
            nn.Linear(num_features, 256),
            nn.ReLU(inplace=True),
            nn.Dropout(0.5),
            # Use the value from the config
            nn.Linear(256, config.num_classes)
        )

    def forward(self, x):
        return self.backbone(x)