Hotfix/alg 1470 drop boxes padding

documentation/source/ObjectDetection.md

@@ -259,12 +259,11 @@ MY_CLASSES = ['cat', 'dog', 'donut']
 @register_dataset("MyNewDetectionDataset")
 class MyNewDetectionDataset(DetectionDataset):
     def __init__(self, data_dir: str, samples_dir: str, targets_dir: str, input_dim: Tuple[int, int],
-                 transforms: List[DetectionTransform], max_targets: int = 100, max_num_samples: int = None, 
+                 transforms: List[DetectionTransform], max_num_samples: int = None, 
                  class_inclusion_list: Optional[List[str]] = None, **kwargs):
         self.sample_paths = None
         self.samples_sub_directory = samples_dir
         self.targets_sub_directory = targets_dir
-        self.max_targets = max_targets
 
         # setting cache as False to be able to load non-resized images and crop in one of transforms
         super().__init__(data_dir=data_dir, input_dim=input_dim,
@@ -310,9 +309,7 @@ class MyNewDetectionDataset(DetectionDataset):
             lines = targets_file.read().splitlines()
             target = np.array([x.strip().strip(',').split(',') for x in lines], dtype=np.float32)
 
-        res_target = np.zeros((self.max_targets, 5))  # cls, cx, cy, w, h
-        if len(target) != 0:
-            res_target[:len(target)] = target
+        res_target = np.array(target) if len(target) != 0 else np.zeros((0, 5))  # cls, cx, cy, w, h
         annotation = {
             'img_path': os.path.join(self.data_dir, sample_path),
             'target': res_target,

src/super_gradients/recipes/dataset_params/coco_detection_dataset_params.yaml

@@ -33,7 +33,6 @@ train_dataset_params:
         prob: 0.5                       # probability to apply horizontal flip
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 120
     - DetectionTargetsFormatTransform:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
         output_format: LABEL_CXCYWH
@@ -65,7 +64,6 @@ val_dataset_params:
   - DetectionPaddedRescale:
       input_dim: ${dataset_params.val_dataset_params.input_dim}
   - DetectionTargetsFormatTransform:
-      max_targets: 50
       input_dim: ${dataset_params.val_dataset_params.input_dim}
       output_format: LABEL_CXCYWH
   tight_box_rotation: False

src/super_gradients/recipes/dataset_params/coco_detection_ppyoloe_dataset_params.yaml

@@ -36,7 +36,6 @@ train_dataset_params:
         mean: [ 123.675, 116.28, 103.53 ]
         std: [ 58.395,  57.12,  57.375 ]
     - DetectionTargetsFormatTransform:
-        max_targets: 256
         output_format: LABEL_CXCYWH
 
   tight_box_rotation: False
@@ -78,7 +77,6 @@ val_dataset_params:
         mean: [ 123.675, 116.28, 103.53 ]
         std: [ 58.395,  57.12,  57.375 ]
     - DetectionTargetsFormatTransform:
-        max_targets: 256
         output_format: LABEL_CXCYWH
   tight_box_rotation: False
   class_inclusion_list:

src/super_gradients/recipes/dataset_params/coco_detection_ssd_lite_mobilenet_v2_dataset_params.yaml

@@ -28,10 +28,8 @@ train_dataset_params:
         prob: 0.5                       # probability to apply horizontal flip
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 120
     - DetectionTargetsFormatTransform:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 50
         output_format: LABEL_NORMALIZED_CXCYWH
 
   tight_box_rotation: False
@@ -63,7 +61,6 @@ val_dataset_params:
         input_dim: ${dataset_params.val_dataset_params.input_dim}
     - DetectionTargetsFormatTransform:
         input_dim: ${dataset_params.val_dataset_params.input_dim}
-        max_targets: 50
         output_format: LABEL_NORMALIZED_CXCYWH
   tight_box_rotation: False
   class_inclusion_list:

src/super_gradients/recipes/dataset_params/coco_detection_yolo_format_base_dataset_params.yaml

@@ -41,7 +41,6 @@ train_dataset_params:
         prob: 0.5                       # probability to apply horizontal flip
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 120
     - DetectionTargetsFormatTransform:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
         output_format: LABEL_CXCYWH
@@ -75,7 +74,6 @@ val_dataset_params:
   - DetectionPaddedRescale:
       input_dim: ${dataset_params.val_dataset_params.input_dim}
   - DetectionTargetsFormatTransform:
-      max_targets: 50
       input_dim: ${dataset_params.val_dataset_params.input_dim}
       output_format: LABEL_CXCYWH
   class_inclusion_list:

src/super_gradients/recipes/dataset_params/coco_detection_yolo_nas_dataset_params.yaml

@@ -32,12 +32,10 @@ train_dataset_params:
         flip_prob: 0.5                  # probability to apply horizontal flip
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 120
         pad_value: 114
     - DetectionStandardize:
         max_value: 255.
     - DetectionTargetsFormatTransform:
-        max_targets: 256
         output_format: LABEL_CXCYWH
 
   tight_box_rotation: False
@@ -71,7 +69,6 @@ val_dataset_params:
         max_value: 255.
     - DetectionImagePermute
     - DetectionTargetsFormatTransform:
-        max_targets: 50
         input_dim: [640, 640]
         output_format: LABEL_CXCYWH
   tight_box_rotation: False

src/super_gradients/recipes/dataset_params/pascal_voc_detection_dataset_params.yaml

@@ -7,7 +7,6 @@ train_dataset_params:
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
     - DetectionTargetsFormatTransform:
-        max_targets: 50
         input_dim: ${dataset_params.train_dataset_params.input_dim}
         output_format: LABEL_CXCYWH
   class_inclusion_list:
@@ -23,7 +22,6 @@ val_dataset_params:
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
     - DetectionTargetsFormatTransform:
-        max_targets: 50
         input_dim: ${dataset_params.train_dataset_params.input_dim}
         output_format: LABEL_CXCYWH
   images_sub_directory: images/test2007/

src/super_gradients/recipes/dataset_params/roboflow_detection_dataset_params.yaml

@@ -39,11 +39,9 @@ train_dataset_params:
         prob: 0.5                       # probability to apply horizontal flip
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
-        max_targets: 300
     - DetectionStandardize:
         max_value: 255.
     - DetectionTargetsFormatTransform:
-        max_targets: 300
         input_dim: ${dataset_params.train_dataset_params.input_dim}
         output_format: LABEL_CXCYWH
   tight_box_rotation: False
@@ -76,12 +74,10 @@ val_dataset_params:
   transforms:
   - DetectionPaddedRescale:
       input_dim: ${dataset_params.val_dataset_params.input_dim}
-      max_targets: 300
       pad_value: 114
   - DetectionStandardize:
       max_value: 255.
   - DetectionTargetsFormatTransform:
-      max_targets: 300
       input_dim: ${dataset_params.val_dataset_params.input_dim}
       output_format: LABEL_CXCYWH
   tight_box_rotation: False

src/super_gradients/training/transforms/transforms.py

@@ -1,6 +1,7 @@
 import collections
 import math
 import random
+import warnings
 from numbers import Number
 from typing import Optional, Union, Tuple, List, Sequence, Dict
 
@@ -776,18 +777,20 @@ class DetectionPaddedRescale(DetectionTransform):
     Preprocessing transform to be applied last of all transforms for validation.
 
     Image- Rescales and pads to self.input_dim.
-    Targets- pads targets to max_targets, moves the class label to first index, converts boxes format- xyxy -> cxcywh.
+    Targets- moves the class label to first index, converts boxes format- xyxy -> cxcywh.
 
     :param input_dim:   Final input dimension (default=(640,640))
     :param swap:        Image axis's to be rearranged.
-    :param max_targets:
     :param pad_value:   Padding value for image.
     """
 
-    def __init__(self, input_dim: Union[int, Tuple[int, int], None], swap: Tuple[int, ...] = (2, 0, 1), max_targets: int = 50, pad_value: int = 114):
+    def __init__(
+        self, input_dim: Union[int, Tuple[int, int], None], swap: Tuple[int, ...] = (2, 0, 1), max_targets: Optional[int] = None, pad_value: int = 114
+    ):
+        super().__init__()
+        _max_targets_deprication(max_targets)
         self.swap = swap
         self.input_dim = ensure_is_tuple_of_two(input_dim)
-        self.max_targets = max_targets
         self.pad_value = pad_value
 
     def __call__(self, sample: dict) -> dict:
@@ -814,20 +817,18 @@ class DetectionHorizontalFlip(DetectionTransform):
     Horizontal Flip for Detection
 
     :param prob:        Probability of applying horizontal flip
-    :param max_targets: Max objects in single image, padding target to this size in case of empty image.
     """
 
-    def __init__(self, prob: float, max_targets: int = 120):
+    def __init__(self, prob: float, max_targets: Optional[int] = None):
         super(DetectionHorizontalFlip, self).__init__()
+        _max_targets_deprication(max_targets)
         self.prob = prob
-        self.max_targets = max_targets
 
     def __call__(self, sample):
         image, targets = sample["image"], sample["target"]
+        if len(targets) == 0:
+            targets = np.zeros((0, 5), dtype=np.float32)
         boxes = targets[:, :4]
-        if len(boxes) == 0:
-            targets = np.zeros((self.max_targets, 5), dtype=np.float32)
-            boxes = targets[:, :4]
         image, boxes = _mirror(image, boxes, self.prob)
         targets[:, :4] = boxes
         sample["target"] = targets
@@ -1012,7 +1013,6 @@ class DetectionTargetsFormatTransform(DetectionTransform):
     :param input_format:       Format of the input targets. For instance [xmin, ymin, xmax, ymax, cls_id] refers to XYXY_LABEL.
     :param output_format:      Format of the output targets. For instance [xmin, ymin, xmax, ymax, cls_id] refers to XYXY_LABEL
     :param min_bbox_edge_size: bboxes with edge size lower then this values will be removed.
-    :param max_targets:        Max objects in single image, padding target to this size.
     """
 
     @resolve_param("input_format", ConcatenatedTensorFormatFactory())
@@ -1023,9 +1023,10 @@ def __init__(
         input_format: ConcatenatedTensorFormat = XYXY_LABEL,
         output_format: ConcatenatedTensorFormat = LABEL_CXCYWH,
         min_bbox_edge_size: float = 1,
-        max_targets: int = 120,
+        max_targets: Optional[int] = None,
     ):
         super(DetectionTargetsFormatTransform, self).__init__()
+        _max_targets_deprication(max_targets)
         if isinstance(input_format, DetectionTargetsFormat) or isinstance(output_format, DetectionTargetsFormat):
             raise TypeError(
                 "DetectionTargetsFormat is not supported for input_format and output_format starting from super_gradients==3.0.7.\n"
@@ -1035,7 +1036,6 @@ def __init__(
             )
         self.input_format = input_format
         self.output_format = output_format
-        self.max_targets = max_targets
         self.min_bbox_edge_size = min_bbox_edge_size
         self.input_dim = None
 
@@ -1066,8 +1066,7 @@ def apply_on_targets(self, targets: np.ndarray) -> np.ndarray:
         """Convert targets in input_format to output_format, filter small bboxes and pad targets"""
         targets = self.targets_format_converter(targets)
         targets = self.filter_small_bboxes(targets)
-        targets = self.pad_targets(targets)
-        return targets
+        return np.ascontiguousarray(targets, dtype=np.float32)
 
     def filter_small_bboxes(self, targets: np.ndarray) -> np.ndarray:
         """Filter bboxes smaller than specified threshold."""
@@ -1078,13 +1077,6 @@ def _is_big_enough(bboxes: np.ndarray) -> np.ndarray:
         targets = filter_on_bboxes(fn=_is_big_enough, tensor=targets, tensor_format=self.output_format)
         return targets
 
-    def pad_targets(self, targets: np.ndarray) -> np.ndarray:
-        """Pad targets."""
-        padded_targets = np.zeros((self.max_targets, targets.shape[-1]))
-        padded_targets[range(len(targets))[: self.max_targets]] = targets[: self.max_targets]
-        padded_targets = np.ascontiguousarray(padded_targets, dtype=np.float32)
-        return padded_targets
-
     def get_equivalent_preprocessing(self) -> List:
         return []
 
@@ -1331,3 +1323,12 @@ def __init__(self, max_val=255.0):
 
     def forward(self, img):
         return img / self.max_val
+
+
+def _max_targets_deprication(max_targets: Optional[int] = None):
+    if max_targets is not None:
+        warnings.warn(
+            "max_targets is deprecated and will be removed in the future, targets are not padded to the max length anymore. "
+            "If you are using collate_fn provided by SG, it is safe to simply drop this argument.",
+            DeprecationWarning,
+        )