[Validate] Add Metadata and Attribute filters to Metrics

.circleci/config.yml

@@ -17,10 +17,11 @@ jobs:
       - run:
           name: Install Environment Dependencies
           command: | # install dependencies
-            apt-get -y install curl
-            pip install --upgrade pip 
+            apt-get update
+            apt-get -y install curl libgeos-dev
+            pip install --upgrade pip
             pip install poetry
-            poetry install
+            poetry install -E shapely
 
       - run:
           name: Black Formatting Check # Only validation, without re-formatting

CHANGELOG.md

@@ -4,6 +4,15 @@ All notable changes to the [Nucleus Python Client](https://github.com/scaleapi/n
 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).
 
+## [0.9.0](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.9.0) - 2022-04-07
+
+### Added
+
+- Validate metrics support metadata and field filtering on input annotation and predictions
+- 3D/Cuboid metrics: Recall, Precision, 3D IOU and birds eye 2D IOU```
+- Shapely can be used for metric development if the optional scale-nucleus[shapely] is installed
+- Full support for passing parameters to evaluation configurations
+
 ## [0.8.4](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.8.4) - 2022-04-06
 - Changing `camera_params` of dataset items can now be done through the dataset method `update_items_metadata`
 

README.md

@@ -179,3 +179,24 @@ cd docs
 sphinx-autobuild . ./_build/html --watch ../nucleus
 ```
 `sphinx-autobuild` will spin up a server on localhost (port 8000 by default) that will watch for and automatically rebuild a version of the API reference based on your local docstring changes.
+
+
+## Custom Metrics using Shapely in scale-validate
+
+Certain metrics use `shapely` which is added as an optional dependency.
+```bash
+pip install scale-nucleus[metrics]
+```
+
+Note that you might need to install a local GEOS package since Shapely doesn't provide binaries bundled with GEOS for every platform.
+
+```bash
+#Mac OS
+brew install geos
+# Ubuntu/Debian flavors
+apt-get install libgeos-dev
+```
+
+To develop it locally use
+
+`poetry install --extra shapely`

nucleus/annotation.py

@@ -557,6 +557,7 @@ class SegmentationAnnotation(Annotation):
     annotations: List[Segment]
     reference_id: str
     annotation_id: Optional[str] = None
+    # metadata: Optional[dict] = None # TODO(sc: 422637)
 
     def __post_init__(self):
         if not self.mask_url:
@@ -574,6 +575,7 @@ def from_json(cls, payload: dict):
             ],
             reference_id=payload[REFERENCE_ID_KEY],
             annotation_id=payload.get(ANNOTATION_ID_KEY, None),
+            # metadata=payload.get(METADATA_KEY, None),  # TODO(sc: 422637)
         )
 
     def to_payload(self) -> dict:
@@ -582,6 +584,7 @@ def to_payload(self) -> dict:
             MASK_URL_KEY: self.mask_url,
             ANNOTATIONS_KEY: [ann.to_payload() for ann in self.annotations],
             ANNOTATION_ID_KEY: self.annotation_id,
+            # METADATA_KEY: self.metadata,  # TODO(sc: 422637)
         }
 
         payload[REFERENCE_ID_KEY] = self.reference_id

nucleus/metrics/__init__.py

@@ -1,5 +1,12 @@
 from .base import Metric, ScalarResult
 from .categorization_metrics import CategorizationF1
+from .cuboid_metrics import CuboidIOU, CuboidPrecision, CuboidRecall
+from .filtering import (
+    FieldFilter,
+    ListOfOrAndFilters,
+    MetadataFilter,
+    apply_filters,
+)
 from .polygon_metrics import (
     PolygonAveragePrecision,
     PolygonIOU,

nucleus/metrics/base.py

@@ -1,9 +1,14 @@
 import sys
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
-from typing import Iterable, List
+from typing import Iterable, List, Optional, Union
 
 from nucleus.annotation import AnnotationList
+from nucleus.metrics.filtering import (
+    ListOfAndFilters,
+    ListOfOrAndFilters,
+    apply_filters,
+)
 from nucleus.prediction import PredictionList
 
 
@@ -86,12 +91,107 @@ def __call__(
         metric(annotations, predictions)
     """
 
+    def __init__(
+        self,
+        annotation_filters: Optional[
+            Union[ListOfOrAndFilters, ListOfAndFilters]
+        ] = None,
+        prediction_filters: Optional[
+            Union[ListOfOrAndFilters, ListOfAndFilters]
+        ] = None,
+    ):
+        """
+        Args:
+            annotation_filters: Filter predicates. Allowed formats are:
+                ListOfAndFilters where each Filter forms a chain of AND predicates.
+                    or
+                ListOfOrAndFilters where Filters are expressed in disjunctive normal form (DNF), like
+                [[MetadataFilter("short_haired", "==", True), FieldFilter("label", "in", ["cat", "dog"]), ...].
+                DNF allows arbitrary boolean logical combinations of single field predicates. The innermost structures
+                each describe a single column predicate. The list of inner predicates is interpreted as a conjunction
+                (AND), forming a more selective `and` multiple field predicate.
+                Finally, the most outer list combines these filters as a disjunction (OR).
+            prediction_filters: Filter predicates. Allowed formats are:
+                ListOfAndFilters where each Filter forms a chain of AND predicates.
+                    or
+                ListOfOrAndFilters where Filters are expressed in disjunctive normal form (DNF), like
+                [[MetadataFilter("short_haired", "==", True), FieldFilter("label", "in", ["cat", "dog"]), ...].
+                DNF allows arbitrary boolean logical combinations of single field predicates. The innermost structures
+                each describe a single column predicate. The list of inner predicates is interpreted as a conjunction
+                (AND), forming a more selective `and` multiple field predicate.
+                Finally, the most outer list combines these filters as a disjunction (OR).
+        """
+        self.annotation_filters = annotation_filters
+        self.prediction_filters = prediction_filters
+
     @abstractmethod
-    def __call__(
+    def call_metric(
         self, annotations: AnnotationList, predictions: PredictionList
     ) -> MetricResult:
         """A metric must override this method and return a metric result, given annotations and predictions."""
 
+    def __call__(
+        self, annotations: AnnotationList, predictions: PredictionList
+    ) -> MetricResult:
+        annotations = self._filter_annotations(annotations)
+        predictions = self._filter_predictions(predictions)
+        return self.call_metric(annotations, predictions)
+
+    def _filter_annotations(self, annotations: AnnotationList):
+        if (
+            self.annotation_filters is None
+            or len(self.annotation_filters) == 0
+        ):
+            return annotations
+        annotations.box_annotations = apply_filters(
+            annotations.box_annotations, self.annotation_filters
+        )
+        annotations.line_annotations = apply_filters(
+            annotations.line_annotations, self.annotation_filters
+        )
+        annotations.polygon_annotations = apply_filters(
+            annotations.polygon_annotations, self.annotation_filters
+        )
+        annotations.cuboid_annotations = apply_filters(
+            annotations.cuboid_annotations, self.annotation_filters
+        )
+        annotations.category_annotations = apply_filters(
+            annotations.category_annotations, self.annotation_filters
+        )
+        annotations.multi_category_annotations = apply_filters(
+            annotations.multi_category_annotations, self.annotation_filters
+        )
+        annotations.segmentation_annotations = apply_filters(
+            annotations.segmentation_annotations, self.annotation_filters
+        )
+        return annotations
+
+    def _filter_predictions(self, predictions: PredictionList):
+        if (
+            self.prediction_filters is None
+            or len(self.prediction_filters) == 0
+        ):
+            return predictions
+        predictions.box_predictions = apply_filters(
+            predictions.box_predictions, self.prediction_filters
+        )
+        predictions.line_predictions = apply_filters(
+            predictions.line_predictions, self.prediction_filters
+        )
+        predictions.polygon_predictions = apply_filters(
+            predictions.polygon_predictions, self.prediction_filters
+        )
+        predictions.cuboid_predictions = apply_filters(
+            predictions.cuboid_predictions, self.prediction_filters
+        )
+        predictions.category_predictions = apply_filters(
+            predictions.category_predictions, self.prediction_filters
+        )
+        predictions.segmentation_predictions = apply_filters(
+            predictions.segmentation_predictions, self.prediction_filters
+        )
+        return predictions
+
     @abstractmethod
     def aggregate_score(self, results: List[MetricResult]) -> ScalarResult:
         """A metric must define how to aggregate results from single items to a single ScalarResult.

nucleus/metrics/categorization_metrics.py

@@ -1,11 +1,12 @@
 from abc import abstractmethod
 from dataclasses import dataclass
-from typing import List, Set, Tuple, Union
+from typing import List, Optional, Set, Tuple, Union
 
 from sklearn.metrics import f1_score
 
 from nucleus.annotation import AnnotationList, CategoryAnnotation
 from nucleus.metrics.base import Metric, MetricResult, ScalarResult
+from nucleus.metrics.filtering import ListOfAndFilters, ListOfOrAndFilters
 from nucleus.metrics.filters import confidence_filter
 from nucleus.prediction import CategoryPrediction, PredictionList
 
@@ -56,12 +57,37 @@ class CategorizationMetric(Metric):
     def __init__(
         self,
         confidence_threshold: float = 0.0,
+        annotation_filters: Optional[
+            Union[ListOfOrAndFilters, ListOfAndFilters]
+        ] = None,
+        prediction_filters: Optional[
+            Union[ListOfOrAndFilters, ListOfAndFilters]
+        ] = None,
     ):
         """Initializes CategorizationMetric abstract object.
 
         Args:
             confidence_threshold: minimum confidence threshold for predictions to be taken into account for evaluation. Must be in [0, 1]. Default 0.0
+            annotation_filters: Filter predicates. Allowed formats are:
+                ListOfAndFilters where each Filter forms a chain of AND predicates.
+                    or
+                ListOfOrAndFilters where Filters are expressed in disjunctive normal form (DNF), like
+                [[MetadataFilter("short_haired", "==", True), FieldFilter("label", "in", ["cat", "dog"]), ...].
+                DNF allows arbitrary boolean logical combinations of single field predicates. The innermost structures
+                each describe a single column predicate. The list of inner predicates is interpreted as a conjunction
+                (AND), forming a more selective `and` multiple field predicate.
+                Finally, the most outer list combines these filters as a disjunction (OR).
+            prediction_filters: Filter predicates. Allowed formats are:
+                ListOfAndFilters where each Filter forms a chain of AND predicates.
+                    or
+                ListOfOrAndFilters where Filters are expressed in disjunctive normal form (DNF), like
+                [[MetadataFilter("short_haired", "==", True), FieldFilter("label", "in", ["cat", "dog"]), ...].
+                DNF allows arbitrary boolean logical combinations of single field predicates. The innermost structures
+                each describe a single column predicate. The list of inner predicates is interpreted as a conjunction
+                (AND), forming a more selective `and` multiple field predicate.
+                Finally, the most outer list combines these filters as a disjunction (OR).
         """
+        super().__init__(annotation_filters, prediction_filters)
         assert 0 <= confidence_threshold <= 1
         self.confidence_threshold = confidence_threshold
 
@@ -83,7 +109,7 @@ def eval(
     def aggregate_score(self, results: List[CategorizationResult]) -> ScalarResult:  # type: ignore[override]
         pass
 
-    def __call__(
+    def call_metric(
         self, annotations: AnnotationList, predictions: PredictionList
     ) -> CategorizationResult:
         if self.confidence_threshold > 0:
@@ -139,7 +165,15 @@ class CategorizationF1(CategorizationMetric):
     """Evaluation method that matches categories and returns a CategorizationF1Result that aggregates to the F1 score"""
 
     def __init__(
-        self, confidence_threshold: float = 0.0, f1_method: str = "macro"
+        self,
+        confidence_threshold: float = 0.0,
+        f1_method: str = "macro",
+        annotation_filters: Optional[
+            Union[ListOfOrAndFilters, ListOfAndFilters]
+        ] = None,
+        prediction_filters: Optional[
+            Union[ListOfOrAndFilters, ListOfAndFilters]
+        ] = None,
     ):
         """
         Args:
@@ -169,8 +203,28 @@ def __init__(
                 Calculate metrics for each instance, and find their average (only
                 meaningful for multilabel classification where this differs from
                 :func:`accuracy_score`).
+            annotation_filters: Filter predicates. Allowed formats are:
+                ListOfAndFilters where each Filter forms a chain of AND predicates.
+                    or
+                ListOfOrAndFilters where Filters are expressed in disjunctive normal form (DNF), like
+                [[MetadataFilter("short_haired", "==", True), FieldFilter("label", "in", ["cat", "dog"]), ...].
+                DNF allows arbitrary boolean logical combinations of single field predicates. The innermost structures
+                each describe a single column predicate. The list of inner predicates is interpreted as a conjunction
+                (AND), forming a more selective `and` multiple field predicate.
+                Finally, the most outer list combines these filters as a disjunction (OR).
+            prediction_filters: Filter predicates. Allowed formats are:
+                ListOfAndFilters where each Filter forms a chain of AND predicates.
+                    or
+                ListOfOrAndFilters where Filters are expressed in disjunctive normal form (DNF), like
+                [[MetadataFilter("short_haired", "==", True), FieldFilter("label", "in", ["cat", "dog"]), ...].
+                DNF allows arbitrary boolean logical combinations of single field predicates. The innermost structures
+                each describe a single column predicate. The list of inner predicates is interpreted as a conjunction
+                (AND), forming a more selective `and` multiple field predicate.
+                Finally, the most outer list combines these filters as a disjunction (OR).
         """
-        super().__init__(confidence_threshold)
+        super().__init__(
+            confidence_threshold, annotation_filters, prediction_filters
+        )
         assert (
             f1_method in F1_METHODS
         ), f"Invalid f1_method {f1_method}, expected one of {F1_METHODS}"