Merge pull request #30 from Vivswan/next

v1.0.5

CHANGELOG.md

@@ -1,5 +1,10 @@
 # Changelog
 
+## 1.0.5 (Patches for Pytorch 2.0.1)
+
+* Removed unnecessary `PseudoParameter.grad` property.
+* Patch for Pytorch 2.0.1, add filtering inputs in `BackwardGraph._calculate_gradients`.
+
 ## 1.0.4
 
 * Combined `PseudoParameter` and `PseudoParameterModule` for better visibility

analogvnn/graph/BackwardGraph.py

@@ -267,7 +267,6 @@ def _calculate_gradients(  # noqa: C901
             grad_inputs = module._call_impl_backward(*grad_outputs.inputs.args, **grad_outputs.inputs.kwargs)
             return ArgsKwargs.to_args_kwargs_object(grad_inputs)
 
-        grad_dict = {}
         inputs = module_inputs_outputs.inputs.args + list(module_inputs_outputs.inputs.kwargs.values())
         outputs = []
         outputs_grads = []
@@ -291,26 +290,28 @@ def _calculate_gradients(  # noqa: C901
             module_parameters = list(module.parameters())
             inputs += module_parameters
 
+        grad_dict = {id(i): None for i in inputs}
+        filtered_inputs = [i for i in inputs if i is not None and i.requires_grad]
         out_grads = torch.autograd.grad(
             outputs=outputs,
-            inputs=inputs,
+            inputs=filtered_inputs,
             grad_outputs=outputs_grads,
             retain_graph=True,
             allow_unused=True
         )
         for i, v in enumerate(out_grads):
-            grad_dict[inputs[i]] = v
+            grad_dict[id(filtered_inputs[i])] = v
 
         for i in module_parameters:
-            if grad_dict[i] is None:
+            if grad_dict[id(i)] is None:
                 continue
 
             if i.grad is None:
-                i.grad = grad_dict[i]
+                i.grad = grad_dict[id(i)]
             else:
-                i.grad += grad_dict[i]
+                i.grad += grad_dict[id(i)]
 
         return ArgsKwargs(
-            args=[grad_dict[i] for i in module_inputs_outputs.inputs.args],
-            kwargs={key: grad_dict[value] for key, value in module_inputs_outputs.inputs.kwargs.items()}
+            args=[grad_dict[id(i)] for i in module_inputs_outputs.inputs.args],
+            kwargs={key: grad_dict[id(value)] for key, value in module_inputs_outputs.inputs.kwargs.items()}
         )

analogvnn/parameter/PseudoParameter.py

@@ -122,26 +122,6 @@ def __repr__(self):
                f', original={self.original}' \
                f')'
 
-    @property
-    def grad(self):
-        """Returns the gradient of the parameter.
-
-        Returns:
-            Tensor: the gradient.
-        """
-
-        return self._transformed.grad
-
-    @grad.setter
-    def grad(self, grad: Tensor):
-        """Sets the gradient of the parameter.
-
-        Args:
-            grad (Tensor): the gradient.
-        """
-
-        self._transformed.grad = grad
-
     @property
     def transformation(self):
         """Returns the transformation.
@@ -187,8 +167,8 @@ def substitute_member(
         """Substitutes a member of a tensor as property of another tensor.
 
         Args:
-            tensor_from (Any): the tensor to substitute from.
-            tensor_to (Any): the tensor to substitute to.
+            tensor_from (Any): the tensor property to substitute.
+            tensor_to (Any): the tensor property to substitute to.
             property_name (str): the name of the property.
             setter (bool): whether to substitute the setter.
         """

pyproject.toml

@@ -14,7 +14,7 @@ py-modules = ['analogvnn']
 [project]
 # $ pip install analogvnn
 name = "analogvnn"
-version = "1.0.4"
+version = "1.0.5"
 description = "A fully modular framework for modeling and optimizing analog/photonic neural networks"  # Optional
 readme = "README.md"
 requires-python = ">=3.7"