Fix torch bug

keras_core/backend/common/variables.py

@@ -14,6 +14,7 @@ class KerasVariable:
         self.name = name or auto_name(self.__class__.__name__)
         dtype = standardize_dtype(dtype)
         self._dtype = dtype
+        self._shape = None
         self._initializer = None
         self.trainable = trainable
         if callable(initializer):

keras_core/backend/torch/layer.py

@@ -7,7 +7,13 @@ class TorchLayer(torch.nn.Module):
     def forward(self, *args, **kwargs):
         # TODO: find a good place to add the params. It should be added right
         # after the variables are initialized.
-        self.params = torch.nn.ParameterList(
-            [variable.value for variable in self.variables]
-        )
+        if not hasattr(self, "torch_params"):
+            self.torch_params = torch.nn.ParameterList(
+                [
+                    torch.nn.Parameter(
+                        variable.value, requires_grad=variable.trainable
+                    )
+                    for variable in self.variables
+                ]
+            )
         return Operation.__call__(self, *args, **kwargs)

keras_core/trainers/data_adapters/py_dataset_adapter.py

@@ -225,7 +225,7 @@ class PyDatasetAdapter(DataAdapter):
                 )
             if len(batch) == 2:
                 sw = data_adapter_utils.class_weight_to_sample_weights(
-                    batch[1], class_weight
+                    batch[1], self.class_weight
                 )
                 batch = batch + (sw,)
         return batch