remove huggingface estimator activation hack

Signed-off-by: Farhan Ahmed <[email protected]>

Author: f4str

art/estimators/classification/hugging_face.py

@@ -324,11 +324,7 @@ def get_activations(  # type: ignore
         def get_feature(name):
             # the hook signature
             def hook(model, input, output):  # pylint: disable=W0622,W0613
-                # TODO: this is using the input, rather than the output, to circumvent the fact
-                # TODO: that flatten is not a layer in pytorch, and the activation defence expects
-                # TODO: a flattened input. A better option is to refactor the activation defence
-                # TODO: to not crash if non 2D inputs are provided.
-                self._features[name] = input
+                self._features[name] = output
 
             return hook