Make cache annotation optional (#1332)

* wip

* remove todos

* Run over whole dataset when cache=True

* fix cache

* fix  test

* improve doc

* update doc

* wip

* fix test

* improve backward compatibility

* update  description

* update  description

* update  description

* undo

src/super_gradients/recipes/dataset_params/coco_detection_dataset_params.yaml

@@ -5,6 +5,8 @@ train_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionMosaic:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
@@ -60,6 +62,8 @@ val_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
   - DetectionPaddedRescale:
       input_dim: ${dataset_params.val_dataset_params.input_dim}

src/super_gradients/recipes/dataset_params/coco_detection_ppyoloe_dataset_params.yaml

@@ -5,6 +5,8 @@ train_dataset_params:
   input_dim: # None, do not resize dataset on load
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionRandomAffine:
         degrees: 0                    # rotation degrees, randomly sampled from [-degrees, degrees]
@@ -70,6 +72,8 @@ val_dataset_params:
   input_dim:
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionRescale:
         output_shape: [640, 640]

src/super_gradients/recipes/dataset_params/coco_detection_ssd_lite_mobilenet_v2_dataset_params.yaml

@@ -8,6 +8,8 @@ train_dataset_params:
   input_dim: [320, 320]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionRandomAffine:
         degrees: 0.                   # rotation degrees, randomly sampled from [-degrees, degrees]
@@ -56,6 +58,8 @@ val_dataset_params:
   input_dim: [320, 320]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionPaddedRescale:
         input_dim: ${dataset_params.val_dataset_params.input_dim}

src/super_gradients/recipes/dataset_params/coco_detection_yolo_format_base_dataset_params.yaml

@@ -13,6 +13,8 @@ train_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionMosaic:
         input_dim: ${dataset_params.train_dataset_params.input_dim}
@@ -70,6 +72,8 @@ val_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
   - DetectionPaddedRescale:
       input_dim: ${dataset_params.val_dataset_params.input_dim}

src/super_gradients/recipes/dataset_params/coco_detection_yolo_nas_dataset_params.yaml

@@ -5,6 +5,8 @@ train_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionRandomAffine:
         degrees: 0                    # rotation degrees, randomly sampled from [-degrees, degrees]
@@ -59,6 +61,8 @@ val_dataset_params:
   input_dim: [636, 636]
   cache_dir:
   cache: False
+  cache_annotations: True
+  ignore_empty_annotations: True
   transforms:
     - DetectionRGB2BGR:
         prob: 1

src/super_gradients/recipes/dataset_params/roboflow_detection_dataset_params.yaml

@@ -9,6 +9,7 @@ train_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
   ignore_empty_annotations: False
   transforms:
     - DetectionMosaic:
@@ -70,6 +71,7 @@ val_dataset_params:
   input_dim: [640, 640]
   cache_dir:
   cache: False
+  cache_annotations: True
   ignore_empty_annotations: False
   transforms:
   - DetectionPaddedRescale:

tests/integration_tests/detection_dataset_test.py

@@ -19,7 +19,7 @@
 
 
 class COCODetectionDataset6Channels(COCODetectionDataset):
-    def get_sample(self, index: int) -> Dict[str, Union[np.ndarray, Any]]:
+    def get_sample(self, index: int, ignore_empty_annotations: bool = False) -> Dict[str, Union[np.ndarray, Any]]:
         img = self.get_resized_image(index)
         img = np.concatenate((img, img), 2)
         annotation = deepcopy(self.annotations[index])

tests/unit_tests/detection_caching.py

@@ -29,8 +29,8 @@ def _load_annotation(self, sample_id: int) -> dict:
         return {"img_path": str(sample_id), "target": np.array([[0, 0, 10, 10, cls_id]]), "resized_img_shape": self.image_size, "seed": sample_id}
 
     # We overwrite this to fake images
-    def _load_image(self, index: int) -> np.ndarray:
-        np.random.seed(self.annotations[index]["seed"])  # Make sure that the generated random tensor of a given index will be the same over the runs
+    def _load_image(self, image_path: str) -> np.ndarray:
+        np.random.seed(int(image_path))
         return np.random.random((self.image_size[0], self.image_size[1], 3)) * 255
 
 

tests/unit_tests/detection_sub_classing_test.py

@@ -32,7 +32,7 @@ def _load_annotation(self, sample_id: int) -> dict:
 
     # DetectionDatasetV2 will call _load_image but since we don't have any image we patch this method with
     # tensor of image shape
-    def _load_image(self, index: int) -> np.ndarray:
+    def _load_image(self, image_path: str) -> np.ndarray:
         return np.random.random(self.image_size)
 
 

tests/unit_tests/detection_sub_sampling_test.py

@@ -25,7 +25,7 @@ def _load_annotation(self, sample_id: int) -> dict:
 
     # DetectionDatasetV2 will call _load_image but since we don't have any image we patch this method with
     # tensor of image shape
-    def _load_image(self, index: int) -> np.ndarray:
+    def _load_image(self, image_path: str) -> np.ndarray:
         return np.random.random(self.image_size)