Started with MEGAN

kgcnn/literature/AttentiveFP/_make.py

@@ -45,6 +45,7 @@
     "dropout": 0.1,
     "verbose": 10,
     "output_embedding": "graph",
+    "output_scaling": None,
     "output_to_tensor": True,  # deprecated
     "output_tensor_type": "padded",
     "output_mlp": {"use_bias": [True, True, False], "units": [25, 10, 1],

kgcnn/literature/GIN/_make.py

@@ -172,7 +172,9 @@ def set_scale(*args, **kwargs):
     "depth": 3, "dropout": 0.0, "verbose": 10,
     "last_mlp": {"use_bias": [True, True, True], "units": [64, 64, 64],
                  "activation": ["relu", "relu", "linear"]},
-    "output_embedding": 'graph', "output_to_tensor": True,
+    "output_embedding": 'graph',
+    "output_to_tensor": None,
+    "output_scaling": None,
     "output_tensor_type": "padded",
     "output_mlp": {"use_bias": True, "units": 1,
                    "activation": "softmax"}

kgcnn/literature/MEGAN/_layers.py

@@ -0,0 +1,31 @@
+from keras import Layer
+from keras import ops
+
+
+class ExplanationSparsityRegularization(Layer):
+
+    def __init__(self,
+                 factor: float = 1.0,
+                 **kwargs):
+        super(ExplanationSparsityRegularization, self).__init__(**kwargs)
+        self.factor = factor
+
+    def build(self, input_shape):
+        super(ExplanationSparsityRegularization, self).build(input_shape)
+
+    def call(self, inputs, **kwargs):
+        r"""Computes a loss from importance scores.
+
+        Args:
+            inputs: Importance tensor of shape ([batch], [N], K) .
+
+        Returns:
+            None.
+        """
+        # importances: ([batch], [N], K)
+        importances = inputs
+
+        loss = ops.mean(ops.abs(importances))
+        loss = loss * self.factor
+        self.add_loss(loss)
+        return loss