Add skip_resize for model.predict

src/super_gradients/training/models/classification_models/base_classifer.py

@@ -30,15 +30,19 @@ def set_dataset_processing_params(self, class_names: Optional[List[str]] = None,
         self._image_processor = image_processor or self._image_processor
 
     @lru_cache(maxsize=1)
-    def _get_pipeline(self, fuse_model: bool = True) -> ClassificationPipeline:
+    def _get_pipeline(self, fuse_model: bool = True, skip_image_resizing: bool = False) -> ClassificationPipeline:
         """Instantiate the prediction pipeline of this model.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
         if None in (self._class_names, self._image_processor):
             raise RuntimeError(
                 "You must set the dataset processing parameters before calling predict.\n" "Please call `model.set_dataset_processing_params(...)` first."
             )
 
+        if skip_image_resizing:
+            raise ValueError("`skip_image_resizing` is not supported for classification models.")
+
         pipeline = ClassificationPipeline(
             model=self,
             image_processor=self._image_processor,
@@ -47,19 +51,21 @@ def _get_pipeline(self, fuse_model: bool = True) -> ClassificationPipeline:
         )
         return pipeline
 
-    def predict(self, images: ImageSource, batch_size: int = 32, fuse_model: bool = True) -> ImagesClassificationPrediction:
+    def predict(self, images: ImageSource, batch_size: int = 32, fuse_model: bool = True, skip_image_resizing: bool = False) -> ImagesClassificationPrediction:
         """Predict an image or a list of images.
 
         :param images:      Images to predict.
         :param batch_size:  Maximum number of images to process at the same time.
         :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(fuse_model=fuse_model)
+        pipeline = self._get_pipeline(fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         return pipeline(images, batch_size=batch_size)  # type: ignore
 
-    def predict_webcam(self, fuse_model: bool = True) -> None:
+    def predict_webcam(self, fuse_model: bool = True, skip_image_resizing: bool = False) -> None:
         """Predict using webcam.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(fuse_model=fuse_model)
+        pipeline = self._get_pipeline(fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         pipeline.predict_webcam()

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

@@ -21,7 +21,7 @@
 import super_gradients.common.factories.detection_modules_factory as det_factory
 from super_gradients.training.utils.predict import ImagesDetectionPrediction
 from super_gradients.training.pipelines.pipelines import DetectionPipeline
-from super_gradients.training.processing.processing import Processing
+from super_gradients.training.processing.processing import Processing, ComposeProcessing, DetectionAutoPadding
 from super_gradients.training.utils.detection_utils import DetectionPostPredictionCallback
 from super_gradients.training.utils.media.image import ImageSource
 
@@ -157,13 +157,16 @@ def get_processing_params(self) -> Optional[Processing]:
         return self._image_processor
 
     @lru_cache(maxsize=1)
-    def _get_pipeline(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True) -> DetectionPipeline:
+    def _get_pipeline(
+        self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True, skip_image_resizing: bool = False
+    ) -> DetectionPipeline:
         """Instantiate the prediction pipeline of this model.
 
         :param iou:     (Optional) IoU threshold for the nms algorithm. If None, the default value associated to the training is used.
         :param conf:    (Optional) Below the confidence threshold, prediction are discarded.
                         If None, the default value associated to the training is used.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
         if None in (self._class_names, self._image_processor, self._default_nms_iou, self._default_nms_conf):
             raise RuntimeError(
@@ -172,9 +175,17 @@ def _get_pipeline(self, iou: Optional[float] = None, conf: Optional[float] = Non
 
         iou = iou or self._default_nms_iou
         conf = conf or self._default_nms_conf
+
+        # Ensure that the image size is divisible by 32.
+        if isinstance(self._image_processor, ComposeProcessing) and skip_image_resizing:
+            image_processor = self._image_processor.get_equivalent_compose_without_resizing()
+            image_processor.processings.append(DetectionAutoPadding(shape_multiple=(32, 32), pad_value=0))
+        else:
+            image_processor = self._image_processor
+
         pipeline = DetectionPipeline(
             model=self,
-            image_processor=self._image_processor,
+            image_processor=image_processor,
             post_prediction_callback=self.get_post_prediction_callback(iou=iou, conf=conf),
             class_names=self._class_names,
             fuse_model=fuse_model,
@@ -188,6 +199,7 @@ def predict(
         conf: Optional[float] = None,
         batch_size: int = 32,
         fuse_model: bool = True,
+        skip_image_resizing: bool = False,
     ) -> ImagesDetectionPrediction:
         """Predict an image or a list of images.
 
@@ -197,19 +209,21 @@ def predict(
                             If None, the default value associated to the training is used.
         :param batch_size:  Maximum number of images to process at the same time.
         :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model)
+        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         return pipeline(images, batch_size=batch_size)  # type: ignore
 
-    def predict_webcam(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True):
+    def predict_webcam(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True, skip_image_resizing: bool = False):
         """Predict using webcam.
 
         :param iou:     (Optional) IoU threshold for the nms algorithm. If None, the default value associated to the training is used.
         :param conf:    (Optional) Below the confidence threshold, prediction are discarded.
                         If None, the default value associated to the training is used.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model)
+        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         pipeline.predict_webcam()
 
     def train(self, mode: bool = True):

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

@@ -20,7 +20,7 @@
 from super_gradients.training.models.detection_models.pp_yolo_e.pp_yolo_head import PPYOLOEHead
 from super_gradients.training.models.sg_module import SgModule
 from super_gradients.training.pipelines.pipelines import DetectionPipeline
-from super_gradients.training.processing.processing import Processing
+from super_gradients.training.processing.processing import Processing, ComposeProcessing, DetectionAutoPadding
 from super_gradients.training.utils import HpmStruct
 from super_gradients.training.utils.media.image import ImageSource
 from super_gradients.training.utils.predict import ImagesDetectionPrediction
@@ -150,13 +150,16 @@ def set_dataset_processing_params(
         self._default_nms_conf = conf or self._default_nms_conf
 
     @lru_cache(maxsize=1)
-    def _get_pipeline(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True) -> DetectionPipeline:
+    def _get_pipeline(
+        self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True, skip_image_resizing: bool = False
+    ) -> DetectionPipeline:
         """Instantiate the prediction pipeline of this model.
 
         :param iou:     (Optional) IoU threshold for the nms algorithm. If None, the default value associated to the training is used.
         :param conf:    (Optional) Below the confidence threshold, prediction are discarded.
                         If None, the default value associated to the training is used.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
         if None in (self._class_names, self._image_processor, self._default_nms_iou, self._default_nms_conf):
             raise RuntimeError(
@@ -166,11 +169,19 @@ def _get_pipeline(self, iou: Optional[float] = None, conf: Optional[float] = Non
         iou = iou or self._default_nms_iou
         conf = conf or self._default_nms_conf
 
+        # Ensure that the image size is divisible by 32.
+        if isinstance(self._image_processor, ComposeProcessing) and skip_image_resizing:
+            image_processor = self._image_processor.get_equivalent_compose_without_resizing()
+            image_processor.processings.append(DetectionAutoPadding(shape_multiple=(32, 32), pad_value=0))
+        else:
+            image_processor = self._image_processor
+
         pipeline = DetectionPipeline(
             model=self,
-            image_processor=self._image_processor,
+            image_processor=image_processor,
             post_prediction_callback=self.get_post_prediction_callback(iou=iou, conf=conf),
             class_names=self._class_names,
+            fuse_model=fuse_model,
         )
         return pipeline
 
@@ -181,6 +192,7 @@ def predict(
         conf: Optional[float] = None,
         batch_size: int = 32,
         fuse_model: bool = True,
+        skip_image_resizing: bool = False,
     ) -> ImagesDetectionPrediction:
         """Predict an image or a list of images.
 
@@ -190,19 +202,21 @@ def predict(
                             If None, the default value associated to the training is used.
         :param batch_size:  Maximum number of images to process at the same time.
         :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model)
+        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         return pipeline(images, batch_size=batch_size)  # type: ignore
 
-    def predict_webcam(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True):
+    def predict_webcam(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True, skip_image_resizing: bool = False):
         """Predict using webcam.
 
         :param iou:     (Optional) IoU threshold for the nms algorithm. If None, the default value associated to the training is used.
         :param conf:    (Optional) Below the confidence threshold, prediction are discarded.
                         If None, the default value associated to the training is used.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model)
+        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         pipeline.predict_webcam()
 
     def train(self, mode: bool = True):

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

@@ -28,7 +28,7 @@
 from super_gradients.training.utils.utils import HpmStruct, check_img_size_divisibility, get_param, infer_model_dtype, infer_model_device
 from super_gradients.training.utils.predict import ImagesDetectionPrediction
 from super_gradients.training.pipelines.pipelines import DetectionPipeline
-from super_gradients.training.processing.processing import Processing
+from super_gradients.training.processing.processing import Processing, ComposeProcessing, DetectionAutoPadding
 from super_gradients.training.utils.media.image import ImageSource
 from super_gradients.module_interfaces import SupportsReplaceInputChannels
 
@@ -536,13 +536,16 @@ def set_dataset_processing_params(
         self._default_nms_conf = conf or self._default_nms_conf
 
     @lru_cache(maxsize=1)
-    def _get_pipeline(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True) -> DetectionPipeline:
+    def _get_pipeline(
+        self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True, skip_image_resizing: bool = False
+    ) -> DetectionPipeline:
         """Instantiate the prediction pipeline of this model.
 
         :param iou:     (Optional) IoU threshold for the nms algorithm. If None, the default value associated to the training is used.
         :param conf:    (Optional) Below the confidence threshold, prediction are discarded.
                         If None, the default value associated to the training is used.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
         if None in (self._class_names, self._image_processor, self._default_nms_iou, self._default_nms_conf):
             raise RuntimeError(
@@ -552,9 +555,16 @@ def _get_pipeline(self, iou: Optional[float] = None, conf: Optional[float] = Non
         iou = iou or self._default_nms_iou
         conf = conf or self._default_nms_conf
 
+        # Ensure that the image size is divisible by 32.
+        if isinstance(self._image_processor, ComposeProcessing) and skip_image_resizing:
+            image_processor = self._image_processor.get_equivalent_compose_without_resizing()
+            image_processor.processings.append(DetectionAutoPadding(shape_multiple=(32, 32), pad_value=0))
+        else:
+            image_processor = self._image_processor
+
         pipeline = DetectionPipeline(
             model=self,
-            image_processor=self._image_processor,
+            image_processor=image_processor,
             post_prediction_callback=self.get_post_prediction_callback(iou=iou, conf=conf),
             class_names=self._class_names,
             fuse_model=fuse_model,
@@ -568,6 +578,7 @@ def predict(
         conf: Optional[float] = None,
         batch_size: int = 32,
         fuse_model: bool = True,
+        skip_image_resizing: bool = False,
     ) -> ImagesDetectionPrediction:
         """Predict an image or a list of images.
 
@@ -577,19 +588,21 @@ def predict(
                             If None, the default value associated to the training is used.
         :param batch_size:  Maximum number of images to process at the same time.
         :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model)
+        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         return pipeline(images, batch_size=batch_size)  # type: ignore
 
-    def predict_webcam(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True):
+    def predict_webcam(self, iou: Optional[float] = None, conf: Optional[float] = None, fuse_model: bool = True, skip_image_resizing: bool = False):
         """Predict using webcam.
 
         :param iou:     (Optional) IoU threshold for the nms algorithm. If None, the default value associated to the training is used.
         :param conf:    (Optional) Below the confidence threshold, prediction are discarded.
                         If None, the default value associated to the training is used.
-        :param fuse_model: If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param fuse_model:  If True, create a copy of the model, and fuse some of its layers to increase performance. This increases memory usage.
+        :param skip_image_resizing: If True, the image processor will not resize the images.
         """
-        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model)
+        pipeline = self._get_pipeline(iou=iou, conf=conf, fuse_model=fuse_model, skip_image_resizing=skip_image_resizing)
         pipeline.predict_webcam()
 
     def train(self, mode: bool = True):