Upload processor

image_processing_swin.py

@@ -0,0 +1,45 @@
+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)
+    
+    image_transform = transforms.Compose([
+        transforms.ToTensor(),
+        # ImageNet normalization
+        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+    ])
+
+    def preprocess(self, images, **kwargs):
+        images = [self.resize_and_pad(image, target_size=224) for image in images]
+        images = [self.image_transform(image) for image in images]
+        images = torch.stack(images)
+        return super().preprocess(images, **kwargs)

preprocessor_config.json

@@ -1,4 +1,7 @@
 {
+  "auto_map": {
+    "AutoImageProcessor": "image_processing_swin.MyCustomSwinProcessor"
+  },
   "do_convert_rgb": null,
   "do_normalize": true,
   "do_rescale": true,