Add way to strip arbitrary output activations (#310)

src/innvestigate/__init__.py

@@ -1,6 +1,9 @@
 from innvestigate import analyzer  # noqa
 from innvestigate.analyzer import create_analyzer  # noqa
 from innvestigate.analyzer.base import NotAnalyzeableModelException  # noqa
-from innvestigate.backend.graph import model_wo_softmax  # noqa
+from innvestigate.backend.graph import (
+    model_wo_softmax,
+    model_wo_output_activation,
+)  # noqa
 
 __version__ = "2.0.1"

src/innvestigate/backend/graph.py

@@ -30,8 +30,9 @@
     "get_layer_neuronwise_io",
     "copy_layer_wo_activation",
     "copy_layer",
-    "pre_softmax_tensors",
+    "pre_output_tensors",
     "model_wo_softmax",
+    "model_wo_output_activation",
     "get_model_layers",
     "model_contains",
     "trace_model_execution",
@@ -338,32 +339,46 @@ def copy_layer(
     return get_layer_from_config(layer, config, weights=weights, **kwargs)
 
 
-def pre_softmax_tensors(Xs: Tensor, should_find_softmax: bool = True) -> list[Tensor]:
-    """Finds the tensors that were preceeding a potential softmax."""
-    softmax_found = False
+def pre_output_tensors(Xs: Tensor, activation: str = None) -> list[Tensor]:
+    """Finds the tensors that were preceeding a potential activation."""
+    activation_found = False
 
     Xs = ibackend.to_list(Xs)
     ret = []
     for x in Xs:
         layer, node_index, _tensor_index = x._keras_history
-        if ichecks.contains_activation(layer, activation="softmax"):
-            softmax_found = True
+        if ichecks.contains_activation(layer, activation=activation):
+            activation_found = True
             if isinstance(layer, klayers.Activation):
                 ret.append(layer.get_input_at(node_index))
             else:
                 layer_wo_act = copy_layer_wo_activation(layer)
                 ret.append(layer_wo_act(layer.get_input_at(node_index)))
 
-    if should_find_softmax and not softmax_found:
-        raise Exception("No softmax found.")
+    if not activation_found:
+        if not activation == None:
+            raise Exception(f"No output activation found.")
+        else:
+            raise Exception(f"No {activation} found.")
 
     return ret
 
 
 def model_wo_softmax(model: Model) -> Model:
     """Creates a new model w/o the final softmax activation."""
     return kmodels.Model(
-        inputs=model.inputs, outputs=pre_softmax_tensors(model.outputs), name=model.name
+        inputs=model.inputs,
+        outputs=pre_output_tensors(model.outputs, activation="softmax"),
+        name=model.name,
+    )
+
+
+def model_wo_output_activation(model: Model) -> Model:
+    """Creates a new model w/o the final activation."""
+    return kmodels.Model(
+        inputs=model.inputs,
+        outputs=pre_output_tensors(model.outputs),
+        name=model.name,
     )