Update to v0.9.1

README.md

@@ -8,7 +8,6 @@
 [![badge-unittests](https://github.com/xaitk/xaitk-saliency/actions/workflows/ci-unittests.yml/badge.svg)](https://github.com/XAITK/xaitk-saliency/actions/workflows/ci-unittests.yml)
 [![badge-notebooks](https://github.com/xaitk/xaitk-saliency/actions/workflows/ci-example-notebooks.yml/badge.svg)](https://github.com/XAITK/xaitk-saliency/actions/workflows/ci-example-notebooks.yml)
 [![codecov](https://codecov.io/gh/XAITK/xaitk-saliency/branch/master/graph/badge.svg?token=VHRNXYCNCG)](https://codecov.io/gh/XAITK/xaitk-saliency)
-[![Language grade: Python](https://img.shields.io/lgtm/grade/python/g/XAITK/xaitk-saliency.svg?logo=lgtm&logoWidth=18)](https://lgtm.com/projects/g/XAITK/xaitk-saliency/context:python)
 
 # XAITK - Saliency
 The `xaitk-saliency` package is an open source, Explainable AI (XAI) framework

docs/release_notes.rst

@@ -17,3 +17,4 @@ Release Notes
    release_notes/v0.8.2
    release_notes/v0.8.3
    release_notes/v0.9.0
+   release_notes/v0.9.1

docs/release_notes/v0.9.1.rst

@@ -0,0 +1,20 @@
+v0.9.1
+======
+
+Fixed a bug where if no detections were found in an image, then the generator would fail.
+
+Updates / New Features
+----------------------
+
+Documentation
+
+* Removed a deprecated badge from the README.
+
+Implementations
+
+* Added a check to exit early if no detections were found in `PerturbationOcclusion`
+
+* Added a check to exit early if no saliency maps were generated in `GenerateObjectDetectorBlackboxSaliency`
+
+Fixes
+-----

pyproject.toml

@@ -8,7 +8,7 @@ name = "xaitk_saliency"
 # REMEMBER: `distutils.version.*Version` types can be used to compare versions
 # from strings like this.
 # This package prefers to use the strict numbering standard when possible.
-version = "0.9.0"
+version = "0.9.1"
 description = """\
     Visual saliency map generation interfaces and baseline implementations \
     for explainable AI."""
@@ -25,7 +25,6 @@ classifiers = [
         'Operating System :: MacOS :: MacOS X',
         'Operating System :: Unix',
         'Programming Language :: Python :: 3',
-        'Programming Language :: Python :: 3.8',
         'Programming Language :: Python :: 3.9',
         'Programming Language :: Python :: 3.10',
         'Programming Language :: Python :: 3.11',

tests/impls/gen_object_detector_blackbox_sal/test_occlusion_based.py

@@ -12,6 +12,10 @@
 from xaitk_saliency.utils.masking import occlude_image_batch
 
 
+def _perturb(ref_image: np.ndarray) -> np.ndarray:
+    return np.ones((6, *ref_image.shape[:2]), dtype=bool)
+
+
 class TestPerturbationOcclusion:
 
     def teardown(self) -> None:
@@ -62,70 +66,40 @@ def test_generate_success(self) -> None:
         Test successfully invoking _generate().
         """
 
-        class StubPI (PerturbImage):
-            """
-            Stub perturber that returns masks of ones.
-            """
-
-            def perturb(self, ref_image: np.ndarray) -> np.ndarray:
-                return np.ones((6, *ref_image.shape[:2]), dtype=bool)
-
-            get_config = None  # type: ignore
-
-        class StubGen (GenerateDetectorProposalSaliency):
-            """
-            Stub saliency generator that returns zeros with correct shape.
-            """
-
-            def generate(
-                self,
-                ref_dets: np.ndarray,
-                pert_dets: np.ndarray,
-                pert_masks: np.ndarray
-            ) -> np.ndarray:
-                return np.zeros((ref_dets.shape[0], *pert_masks.shape[1:]), dtype=np.float16)
-
-            get_config = None  # type: ignore
-
-        class StubDetector (DetectImageObjects):
-            """
-            Stub object detector that returns known detections.
-            """
-
-            def detect_objects(
-                self,
-                img_iter: Iterable[np.ndarray]
-            ) -> Iterable[Iterable[Tuple[AxisAlignedBoundingBox, Dict[Hashable, float]]]]:
-                for i, _ in enumerate(img_iter):
-                    # Return different number of detections for each image to
-                    # test padding functinality
-                    yield [(
-                        AxisAlignedBoundingBox((0, 0), (1, 1)),
-                        {'class0': 0.0, 'class1': 0.9}
-                    ) for _ in range(i)]
-
-            get_config = None  # type: ignore
-
-        test_pi = StubPI()
-        test_gen = StubGen()
-        test_detector = StubDetector()
+        def detect_objects(
+            img_iter: Iterable[np.ndarray]
+        ) -> Iterable[Iterable[Tuple[AxisAlignedBoundingBox, Dict[Hashable, float]]]]:
+            for i, _ in enumerate(img_iter):
+                # Return different number of detections for each image to
+                # test padding functinality
+                yield [(
+                    AxisAlignedBoundingBox((0, 0), (1, 1)),
+                    {'class0': 0.0, 'class1': 0.9}
+                ) for _ in range(i)]
 
         test_image = np.ones((64, 64, 3), dtype=np.uint8)
 
         test_bboxes = np.ones((3, 4))
         test_scores = np.ones((3, 2))
 
+        m_perturb = mock.Mock(spec=PerturbImage)
+        m_perturb.return_value = _perturb(test_image)
+        m_gen = mock.Mock(spec=GenerateDetectorProposalSaliency)
+        m_gen.return_value = np.zeros((3, 64, 64))
+        m_detector = mock.Mock(spec=DetectImageObjects)
+        m_detector.detect_objects = detect_objects
+
         # Call with default fill
         with mock.patch(
             'xaitk_saliency.impls.gen_object_detector_blackbox_sal.occlusion_based.occlude_image_batch',
             wraps=occlude_image_batch
         ) as m_occ_img:
-            inst = PerturbationOcclusion(test_pi, test_gen)
+            inst = PerturbationOcclusion(m_perturb, m_gen)
             test_result = inst._generate(
                 test_image,
                 test_bboxes,
                 test_scores,
-                test_detector
+                m_detector,
             )
 
             assert test_result.shape == (3, 64, 64)
@@ -143,13 +117,13 @@ def detect_objects(
             'xaitk_saliency.impls.gen_object_detector_blackbox_sal.occlusion_based.occlude_image_batch',
             wraps=occlude_image_batch
         ) as m_occ_img:
-            inst = PerturbationOcclusion(test_pi, test_gen)
+            inst = PerturbationOcclusion(m_perturb, m_gen)
             inst.fill = test_fill
             test_result = inst._generate(
                 test_image,
                 test_bboxes,
                 test_scores,
-                test_detector
+                m_detector,
             )
 
             assert test_result.shape == (3, 64, 64)
@@ -160,3 +134,40 @@ def detect_objects(
             m_kwargs = m_occ_img.call_args[-1]
             assert "fill" in m_kwargs
             assert m_kwargs['fill'] == test_fill
+
+    def test_empty_detections(self) -> None:
+        """
+        Test invoking _generate() with empty detections.
+        """
+
+        def detect_objects(
+            img_iter: Iterable[np.ndarray]
+        ) -> Iterable[Iterable[Tuple[AxisAlignedBoundingBox, Dict[Hashable, float]]]]:
+            for i, _ in enumerate(img_iter):
+                # Return 0 detections for each image
+                yield []
+
+        m_detector = mock.Mock(spec=DetectImageObjects)
+        m_detector.detect_objects = detect_objects
+
+        test_image = np.ones((64, 64, 3), dtype=np.uint8)
+
+        test_bboxes = np.ones((3, 4))
+        test_scores = np.ones((3, 2))
+
+        m_perturb = mock.Mock(spec=PerturbImage)
+        m_perturb.return_value = _perturb(test_image)
+        m_gen = mock.Mock(spec=GenerateDetectorProposalSaliency)
+        m_gen.return_value = np.zeros((3, 64, 64))
+        m_detector = mock.Mock(spec=DetectImageObjects)
+        m_detector.detect_objects = detect_objects
+
+        inst = PerturbationOcclusion(m_perturb, m_gen)
+        test_result = inst._generate(
+            test_image,
+            test_bboxes,
+            test_scores,
+            m_detector,
+        )
+
+        assert len(test_result) == 0

tests/interfaces/test_gen_object_detector_blackbox_sal.py

@@ -301,3 +301,38 @@ def test_call_alias() -> None:
         None  # no objectness passed
     )
     assert test_ret == expected_return
+
+
+def test_return_empty_map() -> None:
+    """
+    Test that an empty array of maps is returned properly
+    """
+    m_impl = mock.Mock(spec=GenerateObjectDetectorBlackboxSaliency)
+    m_detector = mock.Mock(spec=DetectImageObjects)
+
+    # test reference detections inputs with matching lengths
+    test_bboxes = np.ones((5, 4), dtype=float)
+    test_scores = np.ones((5, 3), dtype=float)
+
+    # 2-channel image as just HxW should work
+    test_image = np.ones((256, 256), dtype=np.uint8)
+
+    expected_return = np.array([])
+    m_impl._generate.return_value = expected_return
+
+    test_ret = GenerateObjectDetectorBlackboxSaliency.generate(
+        m_impl,
+        test_image,
+        test_bboxes,
+        test_scores,
+        m_detector,
+    )
+
+    m_impl._generate.assert_called_with(
+        test_image,
+        test_bboxes,
+        test_scores,
+        m_detector,
+        None  # no objectness passed
+    )
+    assert len(test_ret) == 0

xaitk_saliency/impls/gen_object_detector_blackbox_sal/occlusion_based.py

@@ -75,6 +75,9 @@ def _generate(
 
         pert_dets_mat = _dets_to_formatted_mat(pert_dets)
 
+        if pert_dets_mat.shape[1] == 0:
+            return np.array([])
+
         return self._generator(
             ref_dets_mat,
             pert_dets_mat,

xaitk_saliency/interfaces/gen_object_detector_blackbox_sal.py

@@ -1,3 +1,5 @@
+import logging
+
 import numpy as np
 import abc
 from typing import Optional
@@ -6,6 +8,7 @@
 from smqtk_detection import DetectImageObjects
 
 from xaitk_saliency.exceptions import ShapeMismatchError
+logger = logging.getLogger(__name__)
 
 
 class GenerateObjectDetectorBlackboxSaliency (Plugfigurable):
@@ -144,6 +147,10 @@ def generate(
             objectness,
         )
 
+        if len(output) == 0:
+            logging.info("No detections found for image. Check DetectImageObjects and saliency configuation")
+            return output
+
         # Check that the saliency heatmaps' shape matches the reference image.
         if output.shape[1:] != ref_image.shape[:2]:
             raise ShapeMismatchError(