fix(losses): fix sce loss buggyness.

Author: okunator

cellseg_models_pytorch/losses/criterions/sce.py

@@ -81,24 +81,15 @@ def forward(
 
         cross_entropy = -torch.sum(forward, dim=1)  # to (B, H, W)
         reverse_cross_entropy = -torch.sum(reverse, dim=1)  # to (B, H, W)
+        loss = self.alpha * cross_entropy + self.beta * reverse_cross_entropy
+
+        if self.apply_sd:
+            loss = self.apply_spectral_decouple(loss, yhat)
 
         if self.class_weights is not None:
-            cross_entropy = self.apply_class_weights(cross_entropy, target)
-            reverse_cross_entropy = self.apply_class_weights(
-                reverse_cross_entropy, target
-            )
+            loss = self.apply_class_weights(loss, target)
 
         if self.edge_weight is not None:
-            cross_entropy = self.apply_edge_weights(cross_entropy, target_weight)
-            reverse_cross_entropy = self.apply_edge_weights(
-                reverse_cross_entropy, target_weight
-            )
-
-        loss = (
-            self.alpha * cross_entropy.mean() + self.beta * reverse_cross_entropy.mean()
-        )
-
-        if self.apply_sd:
-            loss = self.apply_spectral_decouple(loss, yhat)
+            loss = self.apply_edge_weights(loss, target_weight)
 
-        return loss
+        return loss.mean()

changelog.d/20221014_135037_oskari.lehtonen.md

@@ -0,0 +1,3 @@
+## Fixes
+
+- Symmetric CE loss fixed.