Implement dataset.query_objects method (#402)

* Implement dataset.query_objects method

* Remove true_negative from enum

* Fix confusion category true positive case

* Rename IOUMatch to EvaluationMatch

* Rename documentation

* Add documentation to EvaluationMatch

* Propagate model_run_id parameter

* Bump sdk version

---------

Co-authored-by: Gunnar Atli Thoroddsen <[email protected]>

Author: ntamas92

CHANGELOG.md

@@ -6,6 +6,19 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 
+## [0.16.4](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.16.4) - 2023-10-23
+
+### Added
+- Added a `query_objects` method on the Dataset class.
+- Example
+```shell
+>>> ds = client.get_dataset('ds_id')
+>>> objects = ds.query_objects('annotations.metadata.distance_to_device > 150', ObjectQueryType.GROUND_TRUTH_ONLY)
+[CuboidAnnotation(label="", dimensions={}, ...), ...]
+```
+- Added `EvaluationMatch` class to represent IOU Matches, False Positives and False Negatives retrieved through the `query_objects` method 
+
+
 ## [0.16.3](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.16.3) - 2023-10-10
 
 ### Added

nucleus/async_job.py

@@ -177,7 +177,7 @@ def result_urls(self, wait_for_completion=True) -> List[str]:
 
         Parameters:
             wait_for_completion: Defines whether the call shall wait for
-            the job to complete. Defaults to True
+                the job to complete. Defaults to True
 
         Returns:
             A list of signed Scale URLs which contain batches of embeddings.

nucleus/constants.py

@@ -59,6 +59,8 @@
 FX_KEY = "fx"
 FY_KEY = "fy"
 GEOMETRY_KEY = "geometry"
+GROUND_TRUTH_ANNOTATION_ID_KEY = "ground_truth_annotation_id"
+GROUND_TRUTH_ANNOTATION_LABEL_KEY = "ground_truth_annotation_label"
 HEADING_KEY = "heading"
 HEIGHT_KEY = "height"
 ID_KEY = "id"
@@ -68,6 +70,7 @@
 IMAGE_URL_KEY = "image_url"
 INDEX_KEY = "index"
 INDEX_CONTINUOUS_ENABLE_KEY = "enable"
+IOU_KEY = "iou"
 ITEMS_KEY = "items"
 ITEM_KEY = "item"
 ITEM_METADATA_SCHEMA_KEY = "item_metadata_schema"
@@ -97,6 +100,8 @@
 MODEL_TAGS_KEY = "tags"
 MODEL_ID_KEY = "model_id"
 MODEL_RUN_ID_KEY = "model_run_id"
+MODEL_PREDICTION_ID_KEY = "model_prediction_id"
+MODEL_PREDICTION_LABEL_KEY = "model_prediction_label"
 NAME_KEY = "name"
 NEW_ITEMS = "new_items"
 NUCLEUS_ENDPOINT = "https://api.scale.com/v1/nucleus"
@@ -135,6 +140,7 @@
 TRACK_REFERENCE_ID_KEY = "track_reference_id"
 TRACK_REFERENCE_IDS_KEY = "track_reference_ids"
 TRACKS_KEY = "tracks"
+TRUE_POSITIVE_KEY = "true_positive"
 TYPE_KEY = "type"
 UPDATED_ITEMS = "updated_items"
 UPDATE_KEY = "update"

nucleus/dataset.py

@@ -1,5 +1,6 @@
 import datetime
 import os
+from enum import Enum
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -16,7 +17,8 @@
 
 from nucleus.annotation_uploader import AnnotationUploader, PredictionUploader
 from nucleus.async_job import AsyncJob, EmbeddingsExportJob
-from nucleus.prediction import Prediction, from_json
+from nucleus.evaluation_match import EvaluationMatch
+from nucleus.prediction import from_json as prediction_from_json
 from nucleus.track import Track
 from nucleus.url_utils import sanitize_string_args
 from nucleus.utils import (
@@ -77,6 +79,7 @@
     construct_model_run_creation_payload,
     construct_taxonomy_payload,
 )
+from .prediction import Prediction
 from .scene import LidarScene, Scene, VideoScene, check_all_scene_paths_remote
 from .slice import (
     Slice,
@@ -98,6 +101,14 @@
 WARN_FOR_LARGE_SCENES_UPLOAD = 5
 
 
+class ObjectQueryType(str, Enum):
+    IOU = "iou"
+    FALSE_POSITIVE = "false_positive"
+    FALSE_NEGATIVE = "false_negative"
+    PREDICTIONS_ONLY = "predictions_only"
+    GROUND_TRUTH_ONLY = "ground_truth_only"
+
+
 class Dataset:
     """Datasets are collections of your data that can be associated with models.
 
@@ -1681,7 +1692,7 @@ def upload_predictions(
         :class:`Category<CategoryPrediction>`, and :class:`Category<SceneCategoryPrediction>` predictions. Cuboid predictions
         can only be uploaded to a :class:`pointcloud DatasetItem<LidarScene>`.
 
-        When uploading an prediction, you need to specify which item you are
+        When uploading a prediction, you need to specify which item you are
         annotating via the reference_id you provided when uploading the image
         or pointcloud.
 
@@ -1854,7 +1865,7 @@ def prediction_loc(self, model, reference_id, annotation_id):
                 :class:`KeypointsPrediction` \
             ]: Model prediction object with the specified annotation ID.
         """
-        return from_json(
+        return prediction_from_json(
             self._client.make_request(
                 payload=None,
                 route=f"dataset/{self.id}/model/{model.id}/loc/{reference_id}/{annotation_id}",
@@ -1999,6 +2010,47 @@ def query_scenes(self, query: str) -> Iterable[Scene]:
         for item_json in json_generator:
             yield Scene.from_json(item_json, None, True)
 
+    def query_objects(
+        self,
+        query: str,
+        query_type: ObjectQueryType,
+        model_run_id: Optional[str] = None,
+    ) -> Iterable[Union[Annotation, Prediction, EvaluationMatch]]:
+        """
+        Fetches all objects in the dataset that pertain to a given structured query.
+        The results are either Predictions, Annotations, or Evaluation Matches, based on the objectType input parameter
+
+        Args:
+            query: Structured query compatible with the `Nucleus query language <https://nucleus.scale.com/docs/query-language-reference>`_.
+            objectType: Defines the type of the object to query
+
+        Returns:
+            An iterable of either Predictions, Annotations, or Evaluation Matches
+        """
+        json_generator = paginate_generator(
+            client=self._client,
+            endpoint=f"dataset/{self.id}/queryObjectsPage",
+            result_key=ITEMS_KEY,
+            page_size=MAX_ES_PAGE_SIZE,
+            query=query,
+            patch_mode=query_type,
+            model_run_id=model_run_id,
+        )
+
+        for item_json in json_generator:
+            if query_type == ObjectQueryType.GROUND_TRUTH_ONLY:
+                yield Annotation.from_json(item_json)
+            elif query_type == ObjectQueryType.PREDICTIONS_ONLY:
+                yield prediction_from_json(item_json)
+            elif query_type in [
+                ObjectQueryType.IOU,
+                ObjectQueryType.FALSE_POSITIVE,
+                ObjectQueryType.FALSE_NEGATIVE,
+            ]:
+                yield EvaluationMatch.from_json(item_json)
+            else:
+                raise ValueError("Unknown object type", query_type)
+
     @property
     def tracks(self) -> List[Track]:
         """Tracks unique to this dataset.

nucleus/evaluation_match.py

@@ -0,0 +1,102 @@
+from dataclasses import dataclass
+from enum import Enum
+from typing import Optional
+
+from .constants import (
+    DATASET_ITEM_ID_KEY,
+    GROUND_TRUTH_ANNOTATION_ID_KEY,
+    GROUND_TRUTH_ANNOTATION_LABEL_KEY,
+    IOU_KEY,
+    MODEL_PREDICTION_ID_KEY,
+    MODEL_PREDICTION_LABEL_KEY,
+    MODEL_RUN_ID_KEY,
+    TRUE_POSITIVE_KEY,
+)
+
+
+class ConfusionCategory(Enum):
+    TRUE_POSITIVE = "true_positive"
+    FALSE_POSITIVE = "false_positive"
+    FALSE_NEGATIVE = "false_negative"
+
+
+def infer_confusion_category(
+    true_positive: bool,
+    ground_truth_annotation_label: str,
+    model_prediction_label: str,
+):
+    confusion_category = ConfusionCategory.FALSE_NEGATIVE
+
+    if (
+        true_positive
+        or model_prediction_label == ground_truth_annotation_label
+    ):
+        confusion_category = ConfusionCategory.TRUE_POSITIVE
+    elif model_prediction_label is not None:
+        confusion_category = ConfusionCategory.FALSE_POSITIVE
+
+    return confusion_category
+
+
+@dataclass
+class EvaluationMatch:
+    """
+    EvaluationMatch is a result from a model run evaluation. It can represent a true positive, false positive,
+    or false negative.
+
+    The matching only matches the strongest prediction for each annotation, so if there are multiple predictions
+    that overlap a single annotation only the one with the highest overlap metric will be matched.
+
+    The model prediction label and the ground truth annotation label can differ for true positives if there is configured
+    an allowed_label_mapping for the model run.
+
+    NOTE: There is no iou thresholding applied to these matches, so it is possible to have a true positive with a low
+    iou score. If manually rejecting matches remember that a rejected match produces both a false positive and a false
+    negative otherwise you'll skew your aggregates.
+
+    Attributes:
+        model_run_id (str): The ID of the model run that produced this match.
+        model_prediction_id (str): The ID of the model prediction that was matched. None if the match was a false negative.
+        ground_truth_annotation_id (str): The ID of the ground truth annotation that was matched. None if the match was a false positive.
+        iou (int): The intersection over union score of the match.
+        dataset_item_id (str): The ID of the dataset item that was matched.
+        confusion_category (ConfusionCategory): The confusion category of the match.
+        model_prediction_label (str): The label of the model prediction that was matched. None if the match was a false negative.
+        ground_truth_annotation_label (str): The label of the ground truth annotation that was matched. None if the match was a false positive.
+    """
+
+    model_run_id: str
+    model_prediction_id: Optional[str]  # field is nullable
+    ground_truth_annotation_id: Optional[str]  # field is nullable
+    iou: float
+    dataset_item_id: str
+    confusion_category: ConfusionCategory
+    model_prediction_label: Optional[str]  # field is nullable
+    ground_truth_annotation_label: Optional[str]  # field is nullable
+
+    @classmethod
+    def from_json(cls, payload: dict):
+        is_true_positive = payload.get(TRUE_POSITIVE_KEY, False)
+        model_prediction_label = payload.get(MODEL_PREDICTION_LABEL_KEY, None)
+        ground_truth_annotation_label = payload.get(
+            GROUND_TRUTH_ANNOTATION_LABEL_KEY, None
+        )
+
+        confusion_category = infer_confusion_category(
+            true_positive=is_true_positive,
+            ground_truth_annotation_label=ground_truth_annotation_label,
+            model_prediction_label=model_prediction_label,
+        )
+
+        return cls(
+            model_run_id=payload[MODEL_RUN_ID_KEY],
+            model_prediction_id=payload.get(MODEL_PREDICTION_ID_KEY, None),
+            ground_truth_annotation_id=payload.get(
+                GROUND_TRUTH_ANNOTATION_ID_KEY, None
+            ),
+            iou=payload[IOU_KEY],
+            dataset_item_id=payload[DATASET_ITEM_ID_KEY],
+            confusion_category=confusion_category,
+            model_prediction_label=model_prediction_label,
+            ground_truth_annotation_label=ground_truth_annotation_label,
+        )

pyproject.toml

@@ -25,7 +25,7 @@ ignore = ["E501", "E741", "E731", "F401"]  # Easy ignore for getting it running
 
 [tool.poetry]
 name = "scale-nucleus"
-version = "0.16.3"
+version = "0.16.4"
 description = "The official Python client library for Nucleus, the Data Platform for AI"
 license =  "MIT"
 authors = ["Scale AI Nucleus Team <[email protected]>"]