Native LDF Parser (#218)

luxonis_ml/data/parsers/base_parser.py

@@ -5,9 +5,8 @@
 
 from luxonis_ml.data import BaseDataset, DatasetIterator
 from luxonis_ml.enums.enums import DatasetType
-from luxonis_ml.typing import PathType
 
-ParserOutput = Tuple[DatasetIterator, List[str], Dict[str, Dict], List[str]]
+ParserOutput = Tuple[DatasetIterator, Dict[str, Dict], List[Path]]
 """Type alias for parser output.
 
 Contains a function to create the annotation generator, list of classes
@@ -81,7 +80,7 @@ def from_split(self, **kwargs) -> ParserOutput:
         """
         ...
 
-    def _parse_split(self, **kwargs) -> List[str]:
+    def _parse_split(self, **kwargs) -> List[Path]:
         """Parses data in a split subdirectory.
 
         @type kwargs: Dict[str, Any]
@@ -90,7 +89,7 @@ def _parse_split(self, **kwargs) -> List[str]:
         @rtype: List[str]
         @return: List of added images.
         """
-        generator, _, skeletons, added_images = self.from_split(**kwargs)
+        generator, skeletons, added_images = self.from_split(**kwargs)
         self.dataset.add(self._add_task(generator))
         if skeletons:
             for skeleton in skeletons.values():
@@ -151,7 +150,7 @@ def parse_dir(self, dataset_dir: Path, **kwargs) -> BaseDataset:
         return self.dataset
 
     @staticmethod
-    def _get_added_images(generator: DatasetIterator) -> List[PathType]:
+    def _get_added_images(generator: DatasetIterator) -> List[Path]:
         """Returns list of unique images added by the generator
         function.
 
@@ -162,7 +161,7 @@ def _get_added_images(generator: DatasetIterator) -> List[PathType]:
         """
         return list(
             set(
-                item["file"] if isinstance(item, dict) else item.file
+                Path(item["file"] if isinstance(item, dict) else item.file)
                 for item in generator
             )
         )

luxonis_ml/data/parsers/classification_directory_parser.py

@@ -55,7 +55,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(class_dir=dataset_dir / "train")
         added_val_imgs = self._parse_split(class_dir=dataset_dir / "valid")
         added_test_imgs = self._parse_split(class_dir=dataset_dir / "test")
@@ -83,4 +83,4 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), class_names, {}, added_images
+        return generator(), {}, added_images

luxonis_ml/data/parsers/coco_parser.py

@@ -106,7 +106,7 @@ def from_dir(
         use_keypoint_ann: bool = False,
         keypoint_ann_paths: Optional[Dict[str, str]] = None,
         split_val_to_test: bool = True,
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         dir_format, splits = COCOParser._detect_dataset_dir_format(dataset_dir)
         if dir_format is None:
             raise ValueError("Dataset is not in any expected format.")
@@ -208,7 +208,6 @@ def from_split(
         coco_categories = annotation_data["categories"]
         categories = {cat["id"]: cat["name"] for cat in coco_categories}
 
-        class_names = list(categories.values())
         skeletons = {}
         for cat in coco_categories:
             if "keypoints" in cat.keys() and "skeleton" in cat.keys():
@@ -311,7 +310,7 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), class_names, skeletons, added_images
+        return generator(), skeletons, added_images
 
 
 def clean_annotations(annotation_path: Path) -> Path:

luxonis_ml/data/parsers/create_ml_parser.py

@@ -50,7 +50,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "train",
             annotation_path=dataset_dir
@@ -89,7 +89,6 @@ def from_split(
         with open(annotation_path) as f:
             annotations_data = json.load(f)
 
-        class_names = set()
         images_annotations = []
         for annotations in annotations_data:
             path = image_dir.absolute().resolve() / annotations["image"]
@@ -103,7 +102,6 @@ def from_split(
             for curr_ann in annotations["annotations"]:
                 class_name = curr_ann["label"]
                 curr_annotations["classes"].append(class_name)
-                class_names.add(class_name)
 
                 bbox_ann = curr_ann["coordinates"]
                 bbox_xywh = [
@@ -134,4 +132,4 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), list(class_names), {}, added_images
+        return generator(), {}, added_images

luxonis_ml/data/parsers/darknet_parser.py

@@ -49,7 +49,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "train",
             classes_path=dataset_dir / "train" / "_darknet.labels",
@@ -109,4 +109,4 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), list(class_names.values()), {}, added_images
+        return generator(), {}, added_images

luxonis_ml/data/parsers/luxonis_parser.py

@@ -24,6 +24,7 @@
 from .coco_parser import COCOParser
 from .create_ml_parser import CreateMLParser
 from .darknet_parser import DarknetParser
+from .native_parser import NativeParser
 from .segmentation_mask_directory_parser import SegmentationMaskDirectoryParser
 from .solo_parser import SOLOParser
 from .tensorflow_csv_parser import TensorflowCSVParser
@@ -54,6 +55,7 @@ class LuxonisParser(Generic[T]):
         DatasetType.CLSDIR: ClassificationDirectoryParser,
         DatasetType.SEGMASK: SegmentationMaskDirectoryParser,
         DatasetType.SOLO: SOLOParser,
+        DatasetType.NATIVE: NativeParser,
     }
 
     def __init__(
@@ -142,7 +144,8 @@ def __init__(
         dataset_name = dataset_name or name.replace(" ", "_").split(".")[0]
 
         self.dataset = self.dataset_constructor(
-            dataset_name=dataset_name, **kwargs
+            dataset_name=dataset_name,  # type: ignore
+            **kwargs,
         )
         self.parser = self.parsers[self.dataset_type](
             self.dataset, self.dataset_type, task_name

luxonis_ml/data/parsers/native_parser.py

@@ -0,0 +1,73 @@
+import json
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+from luxonis_ml.data import DatasetIterator
+
+from .base_parser import BaseParser, ParserOutput
+
+
+class NativeParser(BaseParser):
+    """Parses directory with native LDF annotations.
+
+    Expected format::
+
+        dataset_dir/
+        ├── train/
+        │   └── annotations.json
+        ├── valid/
+        └── test/
+
+    The annotations are stored in a single JSON file as a list of dictionaries
+    in the same format as the output of the generator function used
+    in L{BaseDataset.add} method.
+    """
+
+    @staticmethod
+    def validate_split(split_path: Path) -> Optional[Dict[str, Any]]:
+        annotation_path = split_path / "annotations.json"
+        if not annotation_path.exists():
+            return None
+        return {"annotation_path": annotation_path}
+
+    @staticmethod
+    def validate(dataset_dir: Path) -> bool:
+        for split in ["train", "valid", "test"]:
+            split_path = dataset_dir / split
+            if NativeParser.validate_split(split_path) is None:
+                return False
+        return True
+
+    def from_dir(
+        self, dataset_dir: Path
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
+        added_train_imgs = self._parse_split(
+            image_dir=dataset_dir / "train",
+            annotation_dir=dataset_dir / "train",
+        )
+        added_val_imgs = self._parse_split(
+            image_dir=dataset_dir / "valid",
+            annotation_dir=dataset_dir / "valid",
+        )
+        added_test_imgs = self._parse_split(
+            image_dir=dataset_dir / "test",
+            annotation_dir=dataset_dir / "test",
+        )
+        return added_train_imgs, added_val_imgs, added_test_imgs
+
+    def from_split(self, annotation_path: Path) -> ParserOutput:
+        """Parses annotations from LDF Format.
+
+        @type annotation_path: C{Path}
+        @param annotation_dir: Path to the JSON file with annotations.
+        @rtype: L{ParserOutput}
+        @return: Annotation generator, list of classes names, skeleton
+            dictionary for keypoints and list of added images.
+        """
+
+        def generator() -> DatasetIterator:
+            yield from json.loads(annotation_path.read_text())
+
+        added_images = self._get_added_images(generator())
+
+        return generator(), {}, added_images

luxonis_ml/data/parsers/segmentation_mask_directory_parser.py

@@ -60,7 +60,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "train",
             seg_dir=dataset_dir / "train",
@@ -96,7 +96,8 @@ def from_split(
             dictionary for keypoints and list of added images
         """
 
-        idx_class = " Class"  # NOTE: space prefix included
+        # NOTE: space prefix included
+        idx_class = " Class"
 
         df = pl.read_csv(classes_path).filter(pl.col(idx_class).is_not_null())
         class_names = df[idx_class].to_list()
@@ -122,4 +123,4 @@ def generator() -> DatasetIterator:
                     }
 
         added_images = self._get_added_images(generator())
-        return generator(), class_names, {}, added_images
+        return generator(), {}, added_images

luxonis_ml/data/parsers/solo_parser.py

@@ -87,12 +87,12 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         """Parses all present data to L{LuxonisDataset} format.
 
         @type dataset_dir: str
         @param dataset_dir: Path to source dataset directory.
-        @rtype: Tuple[List[str], List[str], List[str]]
+        @rtype: Tuple[List[Path], List[Path], List[Path]]
         @return: Tuple with added images for train, valid and test
             splits.
         """
@@ -138,7 +138,7 @@ def from_split(self, split_path: Path) -> ParserOutput:
         )
         # TODO: We make an assumption here that bbox class_names are also valid for all other annotation types in the dataset. Is this OK?
         # TODO: Can we imagine a case where classes between annotation types are different? Which class names to return in this case?
-        if class_names == []:
+        if not class_names:
             raise Exception("No class_names identified. ")
 
         keypoint_labels = self._get_solo_keypoint_names(
@@ -256,7 +256,7 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), class_names, skeletons, added_images
+        return generator(), skeletons, added_images
 
     def _get_solo_annotation_types(
         self, annotation_definitions_dict: dict

luxonis_ml/data/parsers/tensorflow_csv_parser.py

@@ -50,7 +50,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "train",
             annotation_path=dataset_dir / "train" / "_annotations.csv",
@@ -84,7 +84,6 @@ def from_split(
         )
         images_annotations = {}
 
-        class_names = set(df["class"])
         for row in df.rows(named=True):
             path = str(image_dir / str(row["filename"]))
             if path not in images_annotations:
@@ -129,4 +128,4 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), list(class_names), {}, added_images
+        return generator(), {}, added_images

luxonis_ml/data/parsers/voc_parser.py

@@ -47,7 +47,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "train",
             annotation_dir=dataset_dir / "train",
@@ -77,7 +77,6 @@ def from_split(
             dictionary for keypoints and list of added images.
         """
 
-        class_names = set()
         images_annotations = []
         for anno_xml in annotation_dir.glob("*.xml"):
             annotation_data = ET.parse(anno_xml)
@@ -98,7 +97,6 @@ def from_split(
             for object_item in root.findall("object"):
                 class_name = self._xml_find(object_item, "name")
                 curr_annotations["classes"].append(class_name)
-                class_names.add(class_name)
 
                 bbox_info = object_item.find("bndbox")
                 if bbox_info is not None:
@@ -137,7 +135,7 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), list(class_names), {}, added_images
+        return generator(), {}, added_images
 
     @staticmethod
     def _xml_find(root: ET.Element, tag: str) -> str:

luxonis_ml/data/parsers/yolov4_parser.py

@@ -51,7 +51,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[List[str], List[str], List[str]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "train",
             annotation_path=dataset_dir / "train" / "_annotations.txt",
@@ -127,4 +127,4 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), list(class_names.values()), {}, added_images
+        return generator(), {}, added_images

luxonis_ml/data/parsers/yolov6_parser.py

@@ -68,7 +68,7 @@ def validate(dataset_dir: Path) -> bool:
 
     def from_dir(
         self, dataset_dir: Path
-    ) -> Tuple[Optional[List[str]], Optional[List[str]], Optional[List[str]]]:
+    ) -> Tuple[List[Path], List[Path], List[Path]]:
         classes_path = dataset_dir / "data.yaml"
         added_train_imgs = self._parse_split(
             image_dir=dataset_dir / "images" / "train",
@@ -138,4 +138,4 @@ def generator() -> DatasetIterator:
 
         added_images = self._get_added_images(generator())
 
-        return generator(), list(class_names.values()), {}, added_images
+        return generator(), {}, added_images

luxonis_ml/enums/enums.py

@@ -13,3 +13,4 @@ class DatasetType(str, Enum):
     CLSDIR = "clsdir"
     SEGMASK = "segmask"
     SOLO = "solo"
+    NATIVE = "native"