Rename split_channels into attr_dim_summation and invert the logic

Author: dzenanz

captum/attr/_core/layer/grad_cam.py

@@ -82,7 +82,7 @@ def attribute(
         additional_forward_args: Any = None,
         attribute_to_layer_input: bool = False,
         relu_attributions: bool = False,
-        split_channels: bool = False,
+        attr_dim_summation: bool = True,
     ) -> Union[Tensor, Tuple[Tensor, ...]]:
         r"""
         Args:
@@ -150,10 +150,10 @@ def attribute(
                         otherwise, by default, both positive and negative
                         attributions are returned.
                         Default: False
-            split_channels (bool, optional): Indicates whether to
-                        keep attributions split per channel.
-                        The default (False) means to sum per channels.
-                        Default: False
+            attr_dim_summation (bool, optional): Indicates whether to
+                        sum attributions along dimension 1 (usually channel).
+                        The default (True) means to sum along dimension 1.
+                        Default: True
 
         Returns:
             *Tensor* or *tuple[Tensor, ...]* of **attributions**:
@@ -213,16 +213,17 @@ def attribute(
             for layer_grad in layer_gradients
         )
 
-        if split_channels:
+        if attr_dim_summation:
             scaled_acts = tuple(
-                summed_grad * layer_eval
+                torch.sum(summed_grad * layer_eval, dim=1, keepdim=True)
                 for summed_grad, layer_eval in zip(summed_grads, layer_evals)
             )
         else:
             scaled_acts = tuple(
-                torch.sum(summed_grad * layer_eval, dim=1, keepdim=True)
+                summed_grad * layer_eval
                 for summed_grad, layer_eval in zip(summed_grads, layer_evals)
             )
+
         if relu_attributions:
             scaled_acts = tuple(F.relu(scaled_act) for scaled_act in scaled_acts)
         return _format_output(len(scaled_acts) > 1, scaled_acts)

tests/attr/layer/test_grad_cam.py

@@ -47,7 +47,7 @@ def test_simple_input_conv_split_channels(self) -> None:
             net.conv1,
             inp,
             expected_activation=expected_result,
-            split_channels=True,
+            attr_dim_summation=False,
         )
 
     def test_simple_input_conv_no_grad(self) -> None:
@@ -117,7 +117,7 @@ def _grad_cam_test_assert(
         additional_input: Any = None,
         attribute_to_layer_input: bool = False,
         relu_attributions: bool = False,
-        split_channels: bool = False,
+        attr_dim_summation: bool = True,
     ):
         layer_gc = LayerGradCam(model, target_layer)
         self.assertFalse(layer_gc.multiplies_by_inputs)
@@ -127,7 +127,7 @@ def _grad_cam_test_assert(
             additional_forward_args=additional_input,
             attribute_to_layer_input=attribute_to_layer_input,
             relu_attributions=relu_attributions,
-            split_channels=split_channels,
+            attr_dim_summation=attr_dim_summation,
         )
         assertTensorTuplesAlmostEqual(
             self, attributions, expected_activation, delta=0.01