model.py

import torch
import torchvision

from torch import nn
from torchvision import transforms

def create_vit_model(num_classes:int=28, 
                          seed:int=42):
    """Creates an ViTB16 feature extractor model and transforms.

    Args:
        num_classes (int, optional): number of classes in the classifier head. 
            Defaults to 28.
        seed (int, optional): random seed value. Defaults to 42.

    Returns:
        model (torch.nn.Module): ViTB16 feature extractor model. 
        transforms (torchvision.transforms): ViTB16 image transforms.
    """
    # Create ViTB16 pretrained weights, transforms and model
    weights = torchvision.models.ViT_B_16_Weights.DEFAULT

    # Get transforms from weights
    vit_transforms = weights.transforms()

    # Extend the vit_transforms to include grayscale conversion, since vit is trained on 3-channel RGB
    transform_pipeline = transforms.Compose([
        transforms.Grayscale(num_output_channels=3),  # Convert grayscale to 3-channel RGB
        vit_transforms  # Append the existing transforms
    ])
    # transforms = weights.transforms()
    model = torchvision.models.vit_b_16(weights=weights)

    # Freeze all layers in base model
    for param in model.parameters():
        param.requires_grad = False

    # Change heads with random seed for reproducibility
    model.heads = torch.nn.Sequential(
    nn.Linear(in_features=768,
              out_features=28, # Number of Arabic letters = our classes
              bias=True))
    return model, transform_pipeline