feat(utils): add "soft" option to Powerset.to_multilabel conversion (#…

…1516)

CHANGELOG.md

@@ -7,6 +7,7 @@
 - feat(pipeline): add support for list of hooks with `Hooks`
 - BREAKING(pipeline): remove `logging_hook` (use `ArtifactHook` instead)
 - fix(pipeline): add missing "embedding" hook call in `SpeakerDiarization`
+- feat(utils): add `"soft"` option to `Powerset.to_multilabel`
 
 ## Version 3.0.1 (2023-09-28)
 

pyannote/audio/utils/powerset.py

@@ -84,26 +84,32 @@ def build_cardinality(self) -> torch.Tensor:
                 powerset_k += 1
         return cardinality
 
-    def to_multilabel(self, powerset: torch.Tensor) -> torch.Tensor:
-        """Convert predictions from (soft) powerset to (hard) multi-label
+    def to_multilabel(self, powerset: torch.Tensor, soft: bool = False) -> torch.Tensor:
+        """Convert predictions from powerset to multi-label
 
         Parameter
         ---------
         powerset : (batch_size, num_frames, num_powerset_classes) torch.Tensor
             Soft predictions in "powerset" space.
+        soft : bool, optional
+            Return soft multi-label predictions. Defaults to False (i.e. hard predictions)
+            Assumes that `powerset` are "logits" (not "probabilities").
 
         Returns
         -------
         multi_label : (batch_size, num_frames, num_classes) torch.Tensor
-            Hard predictions in "multi-label" space.
+            Predictions in "multi-label" space.
         """
 
-        hard_powerset = torch.nn.functional.one_hot(
-            torch.argmax(powerset, dim=-1),
-            self.num_powerset_classes,
-        ).float()
+        if soft:
+            powerset_probs = torch.exp(powerset)
+        else:
+            powerset_probs = torch.nn.functional.one_hot(
+                torch.argmax(powerset, dim=-1),
+                self.num_powerset_classes,
+            ).float()
 
-        return torch.matmul(hard_powerset, self.mapping)
+        return torch.matmul(powerset_probs, self.mapping)
 
     def forward(self, powerset: torch.Tensor) -> torch.Tensor:
         """Alias for `to_multilabel`"""