model.py

import torch
import torchvision

from torch import nn

def create_vit_model(num_classes:int=3,seed:int=42):
  # Create ViT_B_16 pretrained weights, transforms and model
  weights = torchvision.models.ViT_B_16_Weights.DEFAULT
  transforms = weights.transforms()
  model = torchvision.models.vit_b_16(weights=weights)

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

  # Change classifier head
  torch.manual_seed(seed)
  model.heads = nn.Sequential(nn.Linear(in_features=768, out_features=num_classes))

  return model, transforms