image_processing_swin.py

from torchvision import transforms
from transformers import ViTImageProcessor
import torch
from PIL import Image

class MyCustomSwinProcessor(ViTImageProcessor):
    def resize_and_pad(self, image, target_size=224):
        """Resize image preserving aspect ratio, then pad to target size."""
        # Get original dimensions
        w, h = image.size

        # Calculate scaling factor to fit within target_size while preserving aspect ratio
        scale = min(target_size / w, target_size / h)

        # New dimensions after scaling
        new_w = int(w * scale)
        new_h = int(h * scale)

        # Resize the image
        image = image.resize((new_w, new_h), Image.BILINEAR)

        # Calculate padding needed
        pad_w = target_size - new_w
        pad_h = target_size - new_h

        # Distribute padding evenly on both sides
        left = pad_w // 2
        right = pad_w - left
        top = pad_h // 2
        bottom = pad_h - top

        # Pad with white because its the dataset default background color
        return transforms.functional.pad(image, (left, top, right, bottom), fill=255)

    def preprocess(self, images, **kwargs):
        images = [self.resize_and_pad(image, target_size=224) for image in images]
        images = [transforms.ToTensor()(image) for image in images]
        images = torch.stack(images)
        return super().preprocess(images, **kwargs)