Fix SequentialPreOutputLoss build when loss layer missing

qmarcou · Feb 14, 2024 · 8dab1ed · 8dab1ed
1 parent 5311ba9
commit 8dab1ed
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 1 deletion.
diff --git a/keras_utils/models.py b/keras_utils/models.py
@@ -75,7 +75,7 @@ def build(self, input_shape=None):
         super(SequentialPreOutputLoss, self).build(input_shape=input_shape)
         # In case no Input or batch_shape had been provided output tensors
         # had not been created
-        if self.loss_layer_name is not None and \
+        if self.loss_layer_name is not None and self._loss_layer_added and\
                 self._loss_tensor_output is None:
             self.output_layer_name = self.layers[-1].name
             self._loss_tensor_output = self._loss_layer.output

diff --git a/keras_utils/tests/test_models.py b/keras_utils/tests/test_models.py
@@ -69,6 +69,7 @@ def test_compile(self):
         m.compile()
         m.compile(loss=keras.losses.BinaryCrossentropy(),
                   metrics=keras.metrics.AUC())
+        # Non existing loss layer name
         m = SequentialPreOutputLoss(layers=[inputL, hiddenL, outputL],
                                     loss_layer_name="ZZZ")
         self.assertRaises(ValueError, m.compile)