Fix Semantic segmentation averaging by too many classes (#326)

* Fix Semantic segmentation averaging by too many classes

* Fix FWAVACC

* Set actual version

Author: gatli

CHANGELOG.md

@@ -5,6 +5,16 @@ 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).
 
+## [0.14.5](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.14.5) - 2022-07-05
+
+### Fixed
+- Averaging of rich semantic segmentation taxonomies not taking into account missing classes
+
+## [0.14.4](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.14.4) - 2022-06-21
+
+### Fixed
+- Regression that caused Validate filter statements to not work
+
 ## [0.14.3](https://github.com/scaleapi/nucleus-python-client/releases/tag/v0.14.3) - 2022-06-21
 
 ### Fixed

nucleus/metrics/segmentation_metrics.py

@@ -185,6 +185,19 @@ def _calculate_confusion_matrix(
             ) = convert_to_instance_seg_confusion(
                 confusion, annotation, prediction
             )
+        else:
+            ann_labels = list(
+                dict.fromkeys(s.label for s in annotation.annotations)
+            )
+            pred_labels = list(
+                dict.fromkeys(s.label for s in prediction.annotations)
+            )
+            missing_or_filtered_labels = set(ann_labels) - set(pred_labels)
+            non_taxonomy_classes = {
+                segment.index
+                for segment in annotation.annotations
+                if segment.label in missing_or_filtered_labels
+            }
 
         return confusion, non_taxonomy_classes
 
@@ -644,9 +657,13 @@ def _metric_impl(
                 + confusion.sum(axis=0)
                 - np.diag(confusion)
             )
-            freq = confusion.sum(axis=0) / confusion.sum()
-            fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
-            fwavacc.put(list(non_taxonomy_classes), np.nan)
+            predicted_counts = confusion.sum(axis=0).astype(np.float_)
+            predicted_counts.put(list(non_taxonomy_classes), np.nan)
+            freq = predicted_counts / np.nansum(predicted_counts)
+            iu.put(list(non_taxonomy_classes), np.nan)
+            fwavacc = (
+                np.nan_to_num(freq[freq > 0]) * np.nan_to_num(iu[freq > 0])
+            ).sum()
             mean_fwavacc = np.nanmean(fwavacc)
         return ScalarResult(value=np.nan_to_num(mean_fwavacc), weight=confusion.sum())  # type: ignore
 

pyproject.toml

@@ -21,7 +21,7 @@ exclude = '''
 
 [tool.poetry]
 name = "scale-nucleus"
-version = "0.14.3"
+version = "0.14.5"
 description = "The official Python client library for Nucleus, the Data Platform for AI"
 license =  "MIT"
 authors = ["Scale AI Nucleus Team <nucleusapi@scaleapi.com>"]