Upload processor

image_processing_biomedclip.py

@@ -0,0 +1,26 @@
+from typing import Any, Dict, List, Tuple
+from PIL.Image import Image
+from open_clip.factory import PreprocessCfg, image_transform_v2
+from transformers import ImageProcessingMixin
+import torch
+
+class _ValidKwargs:
+    pass
+
+class BiomedCLIPImageProcessor(ImageProcessingMixin):
+    valid_kwargs = _ValidKwargs
+
+    def __init__(self, preprocess_cfg: Dict[str, Any], 
+                 is_train: bool = False, **kwargs):
+        super().__init__(**kwargs)
+        self.preprocess_cfg = preprocess_cfg
+        self.is_train = is_train
+        self.preprocess = None  # avoid hasattr hack
+        
+    def __call__(self, images, return_tensors="pt"):
+        if not isinstance(images, (list, tuple)):
+            images = [images]
+        if self.preprocess is None:
+            self.preprocess = image_transform_v2(PreprocessCfg(**self.preprocess_cfg), is_train=self.is_train)
+        pixels = torch.stack([self.preprocess(im) for im in images], dim=0)
+        return {"pixel_values": pixels}

preprocessor_config.json

@@ -0,0 +1,28 @@
+{
+  "auto_map": {
+    "AutoImageProcessor": "image_processing_biomedclip.BiomedCLIPImageProcessor"
+  },
+  "image_processor_type": "BiomedCLIPImageProcessor",
+  "is_train": false,
+  "preprocess": null,
+  "preprocess_cfg": {
+    "fill_color": 0,
+    "interpolation": "bicubic",
+    "mean": [
+      0.48145466,
+      0.4578275,
+      0.40821073
+    ],
+    "mode": "RGB",
+    "resize_mode": "shortest",
+    "size": [
+      224,
+      224
+    ],
+    "std": [
+      0.26862954,
+      0.26130258,
+      0.27577711
+    ]
+  }
+}