Bug fix: recall macro average computation when not all classes are pr…

…esent in target and input

Summary: Both f1 and precision are able to perform macro averaging when the input and target both do not contained instances of a class (e.g if num_classes = 5 and both input/target contain only instances of 4 classes). However, recall computation fails in such a case. This diff adds a missing line to make this case work.

Reviewed By: JKSenthil

Differential Revision: D62046138

fbshipit-source-id: 3d6706ed22b8143d61edf8b4891dec6f378b0723

torcheval/metrics/functional/classification/recall.py

@@ -193,6 +193,7 @@ def _recall_compute(
         # Ignore classes which have no samples in `target` and `input`
         mask = (num_labels != 0) | (num_predictions != 0)
         num_tp = num_tp[mask]
+        num_labels = num_labels[mask]
 
     recall = num_tp / num_labels