11import collections
22import math
33import random
4+ import warnings
45from numbers import Number
56from typing import Optional , Union , Tuple , List , Sequence , Dict
67
@@ -776,18 +777,20 @@ class DetectionPaddedRescale(DetectionTransform):
776777 Preprocessing transform to be applied last of all transforms for validation.
777778
778779 Image- Rescales and pads to self.input_dim.
779- Targets- pads targets to max_targets, moves the class label to first index, converts boxes format- xyxy -> cxcywh.
780+ Targets- moves the class label to first index, converts boxes format- xyxy -> cxcywh.
780781
781782 :param input_dim: Final input dimension (default=(640,640))
782783 :param swap: Image axis's to be rearranged.
783- :param max_targets:
784784 :param pad_value: Padding value for image.
785785 """
786786
787- 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 ):
787+ def __init__ (
788+ self , input_dim : Union [int , Tuple [int , int ], None ], swap : Tuple [int , ...] = (2 , 0 , 1 ), max_targets : Optional [int ] = None , pad_value : int = 114
789+ ):
790+ super ().__init__ ()
791+ _max_targets_deprication (max_targets )
788792 self .swap = swap
789793 self .input_dim = ensure_is_tuple_of_two (input_dim )
790- self .max_targets = max_targets
791794 self .pad_value = pad_value
792795
793796 def __call__ (self , sample : dict ) -> dict :
@@ -814,20 +817,18 @@ class DetectionHorizontalFlip(DetectionTransform):
814817 Horizontal Flip for Detection
815818
816819 :param prob: Probability of applying horizontal flip
817- :param max_targets: Max objects in single image, padding target to this size in case of empty image.
818820 """
819821
820- def __init__ (self , prob : float , max_targets : int = 120 ):
822+ def __init__ (self , prob : float , max_targets : Optional [ int ] = None ):
821823 super (DetectionHorizontalFlip , self ).__init__ ()
824+ _max_targets_deprication (max_targets )
822825 self .prob = prob
823- self .max_targets = max_targets
824826
825827 def __call__ (self , sample ):
826828 image , targets = sample ["image" ], sample ["target" ]
829+ if len (targets ) == 0 :
830+ targets = np .zeros ((0 , 5 ), dtype = np .float32 )
827831 boxes = targets [:, :4 ]
828- if len (boxes ) == 0 :
829- targets = np .zeros ((self .max_targets , 5 ), dtype = np .float32 )
830- boxes = targets [:, :4 ]
831832 image , boxes = _mirror (image , boxes , self .prob )
832833 targets [:, :4 ] = boxes
833834 sample ["target" ] = targets
@@ -1012,7 +1013,6 @@ class DetectionTargetsFormatTransform(DetectionTransform):
10121013 :param input_format: Format of the input targets. For instance [xmin, ymin, xmax, ymax, cls_id] refers to XYXY_LABEL.
10131014 :param output_format: Format of the output targets. For instance [xmin, ymin, xmax, ymax, cls_id] refers to XYXY_LABEL
10141015 :param min_bbox_edge_size: bboxes with edge size lower then this values will be removed.
1015- :param max_targets: Max objects in single image, padding target to this size.
10161016 """
10171017
10181018 @resolve_param ("input_format" , ConcatenatedTensorFormatFactory ())
@@ -1023,9 +1023,10 @@ def __init__(
10231023 input_format : ConcatenatedTensorFormat = XYXY_LABEL ,
10241024 output_format : ConcatenatedTensorFormat = LABEL_CXCYWH ,
10251025 min_bbox_edge_size : float = 1 ,
1026- max_targets : int = 120 ,
1026+ max_targets : Optional [ int ] = None ,
10271027 ):
10281028 super (DetectionTargetsFormatTransform , self ).__init__ ()
1029+ _max_targets_deprication (max_targets )
10291030 if isinstance (input_format , DetectionTargetsFormat ) or isinstance (output_format , DetectionTargetsFormat ):
10301031 raise TypeError (
10311032 "DetectionTargetsFormat is not supported for input_format and output_format starting from super_gradients==3.0.7.\n "
@@ -1035,7 +1036,6 @@ def __init__(
10351036 )
10361037 self .input_format = input_format
10371038 self .output_format = output_format
1038- self .max_targets = max_targets
10391039 self .min_bbox_edge_size = min_bbox_edge_size
10401040 self .input_dim = None
10411041
@@ -1066,8 +1066,7 @@ def apply_on_targets(self, targets: np.ndarray) -> np.ndarray:
10661066 """Convert targets in input_format to output_format, filter small bboxes and pad targets"""
10671067 targets = self .targets_format_converter (targets )
10681068 targets = self .filter_small_bboxes (targets )
1069- targets = self .pad_targets (targets )
1070- return targets
1069+ return np .ascontiguousarray (targets , dtype = np .float32 )
10711070
10721071 def filter_small_bboxes (self , targets : np .ndarray ) -> np .ndarray :
10731072 """Filter bboxes smaller than specified threshold."""
@@ -1078,13 +1077,6 @@ def _is_big_enough(bboxes: np.ndarray) -> np.ndarray:
10781077 targets = filter_on_bboxes (fn = _is_big_enough , tensor = targets , tensor_format = self .output_format )
10791078 return targets
10801079
1081- def pad_targets (self , targets : np .ndarray ) -> np .ndarray :
1082- """Pad targets."""
1083- padded_targets = np .zeros ((self .max_targets , targets .shape [- 1 ]))
1084- padded_targets [range (len (targets ))[: self .max_targets ]] = targets [: self .max_targets ]
1085- padded_targets = np .ascontiguousarray (padded_targets , dtype = np .float32 )
1086- return padded_targets
1087-
10881080 def get_equivalent_preprocessing (self ) -> List :
10891081 return []
10901082
@@ -1331,3 +1323,12 @@ def __init__(self, max_val=255.0):
13311323
13321324 def forward (self , img ):
13331325 return img / self .max_val
1326+
1327+
1328+ def _max_targets_deprication (max_targets : Optional [int ] = None ):
1329+ if max_targets is not None :
1330+ warnings .warn (
1331+ "max_targets is deprecated and will be removed in the future, targets are not padded to the max length anymore. "
1332+ "If you are using collate_fn provided by SG, it is safe to simply drop this argument." ,
1333+ DeprecationWarning ,
1334+ )
0 commit comments