Upload processor

image_processing_retinanet.py

@@ -17,11 +17,12 @@ class RetinaNetImageProcessor(BaseImageProcessor):
             model_name: str='retinanet',
             min_size: int=800,
             max_size: int=1333,
+            fixed_size:  Optional[Tuple[int, int]]=None,
             image_mean: list[int]=[0.485, 0.456, 0.406],
             image_std: list[int]=[0.229, 0.224, 0.225],
             topk_candidates: int=1000,
-            nms_thresh=0.5,
-            detections_per_img=300,
+            nms_thresh: float=0.5,
+            detections_per_img: int=300,
             **kwargs
         ):
         super().__init__(**kwargs)
@@ -30,6 +31,7 @@ class RetinaNetImageProcessor(BaseImageProcessor):
         self.config = {
             'min_size': min_size,
             'max_size': max_size,
+            'fixed_size': fixed_size,
             'image_mean': image_mean,
             'image_std': image_std
         }
@@ -42,7 +44,7 @@ class RetinaNetImageProcessor(BaseImageProcessor):
     def post_process_object_detection(
             self,
             outputs,
-            threshold: float = 0.5,
+            threshold: float = 0.05,
             target_sizes: List[Tuple] = None
         ) -> List[Dict[str, torch.Tensor]]:
         """
@@ -54,10 +56,9 @@ class RetinaNetImageProcessor(BaseImageProcessor):
         logits = outputs.logits
         pred_boxes = outputs.pred_boxes
         image_sizes = [(int(image_size[0]), int(image_size[1])) for image_size in outputs.image_sizes]
-        anchors = outputs.anchors
-        features = outputs.features
+        anchors = list(outputs.anchors)
+        num_anchors_per_level = outputs.num_anchors_per_level
 
-        num_anchors_per_level = [x.size(2) * x.size(3) for x in features]
         HW = 0
         for v in num_anchors_per_level:
             HW += v
@@ -142,6 +143,7 @@ class RetinaNetImageProcessor(BaseImageProcessor):
     def preprocess(
             self, 
             images, 
+            fixed_size:  Optional[Tuple[int, int]]=None,
             annotations: Optional[Union[AnnotationType, List[AnnotationType]]] = None,
         ) -> BatchFeature:
         """
@@ -156,6 +158,9 @@ class RetinaNetImageProcessor(BaseImageProcessor):
         if images is not None and not isinstance(images, list):
             images = [images]
         
+        if fixed_size is not None:
+            self.config['fixed_size'] = fixed_size
+
         transform = GeneralizedRCNNTransform(**self.config)
         totensor = ToTensor()
         image_tensors = [totensor(img) for img in images]   

preprocessor_config.json

@@ -3,6 +3,7 @@
     "AutoImageProcessor": "image_processing_retinanet.RetinaNetImageProcessor"
   },
   "config": {
+    "fixed_size": null,
     "image_mean": [
       0.485,
       0.456,