Feature/sg 1442 sliding window inference for yolonas

src/super_gradients/examples/predict/sliding_sindow_detection_predict.py

@@ -0,0 +1,20 @@
+import torch
+from super_gradients.common.object_names import Models
+from super_gradients.training import models
+
+
+# Note that currently only YoloX, PPYoloE and YOLO-NAS are supported.
+from super_gradients.training.models.detection_models.sliding_window_detection_forward_wrapper import SlidingWindowInferenceDetectionWrapper
+
+model = models.get(Models.YOLO_NAS_S, pretrained_weights="coco")
+
+# We want to use cuda if available to speed up inference.
+model = model.to("cuda" if torch.cuda.is_available() else "cpu")
+
+model = SlidingWindowInferenceDetectionWrapper(model=model, tile_size=640, tile_step=160, tile_nms_conf=0.35)
+
+predictions = model.predict(
+    "https://images.pexels.com/photos/7968254/pexels-photo-7968254.jpeg?auto=compress&cs=tinysrgb&w=1260&h=750&dpr=2", skip_image_resizing=True
+)
+predictions.show()
+predictions.save(output_path="2.jpg")  # Save in working directory

src/super_gradients/training/metrics/detection_metrics.py

@@ -33,6 +33,8 @@ class DetectionMetrics(Metric):
 
     :param num_cls:                         Number of classes.
     :param post_prediction_callback:        DetectionPostPredictionCallback to be applied on net's output prior to the metric computation (NMS).
+     When None, the direct outputs of the model will be used.
+
     :param normalize_targets:               Whether to normalize bbox coordinates by image size.
     :param iou_thres:                       IoU threshold to compute the mAP.
                                             Could be either instance of IouThreshold, a tuple (lower bound, upper_bound) or single scalar.
@@ -179,7 +181,8 @@ def update(self, preds, target: torch.Tensor, device: str, inputs: torch.tensor,
         targets = target.clone()
         crowd_targets = torch.zeros(size=(0, 6), device=device) if crowd_targets is None else crowd_targets.clone()
 
-        preds = self.post_prediction_callback(preds, device=device)
+        if self.post_prediction_callback is not None:
+            preds = self.post_prediction_callback(preds, device=device)
 
         new_matching_info = compute_detection_matching(
             preds,

src/super_gradients/training/models/detection_models/customizable_detector.py

@@ -194,9 +194,24 @@ def set_dataset_processing_params(
         if class_agnostic_nms is not None:
             self._default_class_agnostic_nms = bool(class_agnostic_nms)
 
+    def get_dataset_processing_params(self):
+        return dict(
+            class_names=self._class_names,
+            image_processor=self._image_processor,
+            iou=self._default_nms_iou,
+            conf=self._default_nms_iou,
+            nms_top_k=self._default_nms_top_k,
+            max_predictions=self._default_max_predictions,
+            multi_label_per_box=self._default_multi_label_per_box,
+            class_agnostic_nms=self._default_class_agnostic_nms,
+        )
+
     def get_processing_params(self) -> Optional[Processing]:
         return self._image_processor
 
+    def get_class_names(self) -> Optional[List[str]]:
+        return self._class_names
+
     @lru_cache(maxsize=1)
     def _get_pipeline(
         self,