Upload processor

preprocessor_config.json

@@ -1,15 +1,6 @@
 {
-  "crop_pct": 0.875,
-  "do_normalize": true,
-  "do_rescale": false,
-  "do_resize": false,
-  "image_mean": 0.1307,
-  "image_processor_type": "ConvNextImageProcessor",
-  "image_std": 0.3081,
-  "resample": null,
-  "rescale_factor": 0.00392156862745098,
-  "size": {
-    "height": 28,
-    "width": 28
-  }
+  "auto_map": {
+    "AutoImageProcessor": "preprocessor_lenet.LeNetProcessor"
+  },
+  "image_processor_type": "LeNetProcessor"
 }

preprocessor_lenet.py

@@ -0,0 +1,40 @@
+import numpy as np
+from PIL import Image
+from transformers import BaseImageProcessor, BatchFeature
+from transformers.image_utils import ImageInput
+import torch
+from torchvision.transforms import v2
+
+
+class LeNetProcessor(BaseImageProcessor):
+    """
+    A custom processor that only normalizes a grayscale image
+    and prepares it for a model.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(self, **kwargs):
+        """
+        Args:
+        """
+        super().__init__(**kwargs)
+
+    def preprocess(self, images: ImageInput, return_tensors=None, **kwargs) -> BatchFeature:
+        """
+        Preprocess a batch of grayscale images.
+        """
+        if not isinstance(images, list):
+            images = [images]
+
+        transform = v2.Compose([
+            v2.RandomResizedCrop(size=(28, 28), antialias=True),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(
+                mean=[0.1307],
+                std=[0.3081]
+            ),
+        ])
+
+        data = {"pixel_values": transform(images)}
+        return BatchFeature(data=data, tensor_type=return_tensors)