model.py

import torch
import torchvision

from torch import nn

def create_ViT():
  ViT_weights = torchvision.models.ViT_B_16_Weights.DEFAULT
  ViT_model = torchvision.models.vit_b_16(weights=ViT_weights)

  # Freeze pre-trained weights
  for param in ViT_model.parameters():
    param.requires_grad = False

  # Find the encoder module and its layers
  encoder = ViT_model.encoder
  encoder_layers = encoder.layers
  # Modify each Encoder layer to include dropout
  for layer in encoder_layers:
    # Access the Multi-head Self-Attention module (might be named differently)
    attn_module = layer.self_attention  # Replace with the actual module name in your model
    # Add dropout layer after the attention module
    attn_module.add_module('my_dropout', nn.Dropout(p=0.4))
  # Add your new head for classification (same as before)
  ViT_model.heads = nn.Sequential(
      nn.Dropout(p=0.5),
      nn.Linear(in_features=768, out_features=1, bias=True)
  )

  manual_transforms = torchvision.transforms.Compose([
    torchvision.transforms.RandomRotation(25),
    torchvision.transforms.RandomAffine(degrees=0, translate=(0.15, 0.15), shear=15),
    torchvision.transforms.RandomHorizontalFlip(),
    torchvision.transforms.RandomVerticalFlip(),
    torchvision.transforms.ColorJitter(brightness=(0.9, 1.5)),
    torchvision.transforms.Resize((224, 224)),
    torchvision.transforms.ToTensor()
  ])

  return ViT_model, manual_transforms