Fix layer_gradient_x_activation and add logging for metrics (#643)

Summary:
+ Adding logging for captum.metrics
+ currently layer_gradient_x_activation is failing for integer inputs. Adding datatype check before calling `apply_gradient_requirements` otherwise if attribute is called in `torch.no_grad()` context then it fails.

Pull Request resolved: #643

Reviewed By: vivekmig

Differential Revision: D27723943

Pulled By: NarineK

fbshipit-source-id: 96a5fc1a6b848007e59212cbe7f469546fed5178

Author: NarineK

captum/metrics/_core/infidelity.py

@@ -17,6 +17,7 @@
     safe_div,
 )
 from captum._utils.typing import BaselineType, TargetType, TensorOrTupleOfTensorsGeneric
+from captum.log import log_usage
 from captum.metrics._utils.batching import _divide_and_aggregate_metrics
 
 
@@ -108,6 +109,7 @@ def default_perturb_func(
     return sub_infidelity_perturb_func_decorator
 
 
+@log_usage()
 def infidelity(
     forward_func: Callable,
     perturb_func: Callable,

captum/metrics/_core/sensitivity.py

@@ -16,6 +16,7 @@
     _format_tensor_into_tuples,
 )
 from captum._utils.typing import TensorOrTupleOfTensorsGeneric
+from captum.log import log_usage
 from captum.metrics._utils.batching import _divide_and_aggregate_metrics
 
 
@@ -57,6 +58,7 @@ def default_perturb_func(
     return perturbed_input
 
 
+@log_usage()
 def sensitivity_max(
     explanation_func: Callable,
     inputs: TensorOrTupleOfTensorsGeneric,

tests/attr/layer/test_layer_gradient_x_activation.py

@@ -23,6 +23,12 @@ def test_simple_input_gradient_activation(self) -> None:
         inp = torch.tensor([[0.0, 100.0, 0.0]], requires_grad=True)
         self._layer_activation_test_assert(net, net.linear0, inp, [0.0, 400.0, 0.0])
 
+    def test_simple_input_gradient_activation_no_grad(self) -> None:
+        net = BasicModel_MultiLayer()
+        inp = torch.tensor([[0.0, 100.0, 0.0]], requires_grad=True)
+        with torch.no_grad():
+            self._layer_activation_test_assert(net, net.linear0, inp, [0.0, 400.0, 0.0])
+
     def test_simple_linear_gradient_activation(self) -> None:
         net = BasicModel_MultiLayer()
         inp = torch.tensor([[0.0, 100.0, 0.0]])