Merge pull request #139 from twosixlabs/120-explainable-ai-outputs

120 explainable ai outputs

examples/pyproject.toml

@@ -27,7 +27,10 @@ dependencies = [
 cv = [
     # 1.4.1 adds version constraints to scikit-learn and numpy that are incompatible with ART
     "albumentations<1.4.1",
+    "captum",
+    "scikit-learn>=1.2", # required by xaitk-saliency
     "tidecv",
+    "xaitk-saliency",
     "yolov5",
 ]
 

examples/src/armory/examples/image_classification/food101.py

@@ -19,8 +19,10 @@
 import armory.dataset
 import armory.engine
 import armory.evaluation
+import armory.export.captum
 import armory.export.criteria
 import armory.export.image_classification
+import armory.export.xaitksaliency
 import armory.metric
 import armory.metrics.compute
 import armory.metrics.perturbation
@@ -81,7 +83,9 @@ def load_model():
     armory_model = armory.model.image_classification.ImageClassifier(
         name="ViT-finetuned-food101",
         model=hf_model,
-        accessor=armory.data.Images.as_torch(scale=normalized_scale),
+        accessor=armory.data.Images.as_torch(
+            dim=armory.data.ImageDimensions.CHW, scale=normalized_scale
+        ),
     )
 
     art_classifier = armory.track.track_init_params(
@@ -270,12 +274,24 @@ def create_metrics():
     }
 
 
-def create_exporters(export_every_n_batches):
+def create_exporters(model, export_every_n_batches):
     """Create sample exporters"""
     return [
         armory.export.image_classification.ImageClassificationExporter(
             criterion=armory.export.criteria.every_n_batches(export_every_n_batches)
         ),
+        armory.export.captum.CaptumImageClassificationExporter(
+            model,
+            criterion=armory.export.criteria.every_n_batches(export_every_n_batches),
+        ),
+        armory.export.xaitksaliency.XaitkSaliencyBlackboxImageClassificationExporter(
+            name="slidingwindow",
+            model=model,
+            classes=[6, 23],  # beignets(6), churros(23)
+            criterion=armory.export.criteria.when_metric_in(
+                armory.export.criteria.batch_targets(), [6, 23]
+            ),
+        ),
     ]
 
 
@@ -303,7 +319,7 @@ def main(
     )
     perturbations = dict()
     metrics = create_metrics()
-    exporters = create_exporters(export_every_n_batches)
+    exporters = create_exporters(model, export_every_n_batches)
     profiler = armory.metrics.compute.BasicProfiler()
 
     if "benign" in chains:

examples/src/armory/examples/object_detection/coco_fasterrcnn_resnet50.py

@@ -20,6 +20,7 @@
 import armory.engine
 import armory.evaluation
 import armory.export.criteria
+import armory.export.drise
 import armory.export.object_detection
 import armory.metric
 import armory.metrics.compute
@@ -168,12 +169,18 @@ def create_metrics():
     }
 
 
-def create_exporters(export_every_n_batches):
+def create_exporters(model, export_every_n_batches):
     """Create sample exporters"""
     return [
         armory.export.object_detection.ObjectDetectionExporter(
             criterion=armory.export.criteria.every_n_batches(export_every_n_batches)
         ),
+        armory.export.drise.DRiseSaliencyObjectDetectionExporter(
+            model,
+            criterion=armory.export.criteria.every_n_batches(export_every_n_batches),
+            num_classes=91,
+            num_masks=10,
+        ),
     ]
 
 
@@ -188,7 +195,7 @@ def main(batch_size, export_every_n_batches, num_batches, seed, shuffle):
     dataset = load_dataset(batch_size, shuffle)
     attack = create_attack(art_detector, batch_size)
     metrics = create_metrics()
-    exporters = create_exporters(export_every_n_batches)
+    exporters = create_exporters(model, export_every_n_batches)
 
     evaluation = armory.evaluation.Evaluation(
         name="coco-detection-fasterrcnn-resnet50",

examples/src/armory/examples/object_detection/license_plates_yolos.py

@@ -21,6 +21,7 @@
 import armory.engine
 import armory.evaluation
 import armory.export.criteria
+import armory.export.drise
 import armory.export.object_detection
 import armory.metric
 import armory.metrics.compute
@@ -198,12 +199,18 @@ def create_metrics():
     }
 
 
-def create_exporters(export_every_n_batches):
+def create_exporters(model, export_every_n_batches):
     """Create sample exporters"""
     return [
         armory.export.object_detection.ObjectDetectionExporter(
             criterion=armory.export.criteria.every_n_batches(export_every_n_batches)
         ),
+        armory.export.drise.DRiseSaliencyObjectDetectionExporter(
+            model,
+            criterion=armory.export.criteria.every_n_batches(export_every_n_batches),
+            num_classes=2,
+            num_masks=10,
+        ),
     ]
 
 
@@ -218,7 +225,7 @@ def main(batch_size, export_every_n_batches, num_batches, seed, shuffle):
     dataset = load_dataset(batch_size, shuffle)
     attack = create_attack(art_detector, batch_size)
     metrics = create_metrics()
-    exporters = create_exporters(export_every_n_batches)
+    exporters = create_exporters(model, export_every_n_batches)
 
     evaluation = armory.evaluation.Evaluation(
         name="license-plate-detection-yolos",

examples/src/armory/examples/object_detection/license_plates_yolov5.py

@@ -21,6 +21,7 @@
 import armory.engine
 import armory.evaluation
 import armory.export.criteria
+import armory.export.drise
 import armory.export.object_detection
 import armory.metric
 import armory.metrics.compute
@@ -183,12 +184,18 @@ def create_metrics():
     }
 
 
-def create_exporters(export_every_n_batches):
+def create_exporters(model, export_every_n_batches):
     """Create sample exporters"""
     return [
         armory.export.object_detection.ObjectDetectionExporter(
             criterion=armory.export.criteria.every_n_batches(export_every_n_batches)
         ),
+        armory.export.drise.DRiseSaliencyObjectDetectionExporter(
+            model,
+            criterion=armory.export.criteria.every_n_batches(export_every_n_batches),
+            num_classes=1,
+            num_masks=10,
+        ),
     ]
 
 
@@ -203,7 +210,7 @@ def main(batch_size, export_every_n_batches, num_batches, seed, shuffle):
     dataset = load_dataset(batch_size, shuffle)
     attack = create_attack(art_detector, batch_size)
     metrics = create_metrics()
-    exporters = create_exporters(export_every_n_batches)
+    exporters = create_exporters(model, export_every_n_batches)
 
     evaluation = armory.evaluation.Evaluation(
         name="license-plate-detection-yolov5",

examples/src/armory/examples/utils/display.py

@@ -93,7 +93,7 @@ def display_object_detection_results(
         for sample_idx in range(batch_size):
             for chain_idx, chain in enumerate(chains):
                 image_filename = Exporter.artifact_path(
-                    chain, batch_idx, sample_idx, "input.png"
+                    chain, batch_idx, sample_idx, "objects.png"
                 )
                 client.download_artifacts(run_id, image_filename, tmpdir)
                 image = plt.imread(tmppath / image_filename)
@@ -112,6 +112,7 @@ def display_object_detection_results(
                 ax.tick_params(
                     bottom=False, left=False, labelbottom=False, labelleft=False
                 )
+                ax.axis("off")
 
     fig.suptitle(f"Batch {batch_idx}")
     fig.tight_layout()

library/pyproject.toml

@@ -42,10 +42,12 @@ dependencies = [
 extras = [
     "adversarial-robustness-toolbox",
     "albumentations",
+    "captum",
     "matplotlib",
     "pandas",
     "tidecv",
     "transformers",
+    "xaitk-saliency",
 ]