feat: customize image processor in __call__

processing_clip.py

@@ -26,6 +26,14 @@ class JinaCLIPProcessor(CLIPProcessor):
 
 class JinaCLIPImageProcessor(BaseImageProcessor):
     model_input_names = ['pixel_values']
+    _valid_processor_keys = [
+        'size',
+        'mean',
+        'std',
+        'resize_mode',
+        'interpolation',
+        'fill_color',
+    ]
 
     def __init__(
         self,
@@ -44,14 +52,17 @@ class JinaCLIPImageProcessor(BaseImageProcessor):
         self.resize_mode = resize_mode
         self.interpolation = interpolation
         self.fill_color = fill_color
-        self.transform = image_transform(
-            image_size=size,
+        self.transform = self._build_transform()
+
+    def _build_transform(self):
+        return image_transform(
+            image_size=self.size,
             is_train=False,
-            mean=mean,
-            std=std,
-            resize_mode=resize_mode,
-            interpolation=interpolation,
-            fill_color=fill_color,
+            mean=self.mean,
+            std=self.std,
+            resize_mode=self.resize_mode,
+            interpolation=self.interpolation,
+            fill_color=self.fill_color,
             aug_cfg=None,
         )
 
@@ -61,6 +72,17 @@ class JinaCLIPImageProcessor(BaseImageProcessor):
         return output
 
     def preprocess(self, images: ImageInput, **kwargs) -> BatchFeature:
+
+        _transform_needs_rebuild = False
+        for k, v in kwargs.items():
+            if k in self._valid_processor_keys:
+                if v != getattr(self, k):
+                    setattr(self, k, v)
+                    _transform_needs_rebuild = True
+
+        if _transform_needs_rebuild:
+            self.transform = self._build_transform()
+
         images = make_list_of_images(images)
         out = torch.stack([self.transform(image) for image in images], dim=0)
         return BatchFeature(data={'pixel_values': out})