[keras] Importance: fixed grad importance + adds sample_weights

keras.py

@@ -1315,18 +1315,21 @@ class LBNLayer(tf.keras.layers.Layer):
 
 
 def feature_importance_grad(model, x=None, **kwargs):
-    inp = [tf.Variable(v) for v in x]
+    inp = [tf.constant(v) for v in x]
+
     with tf.GradientTape() as tape:
+        tape.watch(inp)
         pred = model(inp, training=False)
         ix = np.argsort(pred, axis=-1)[:, -1]
         decision = tf.gather(pred, ix, batch_dims=1)
-
     gradients = tape.gradient(decision, inp)  # gradients for decision nodes
-    normed_gradients = [_g * _x for (_g, _x) in zip(gradients, x)]  # normed to input values
-
-    mean_gradients = np.concatenate(
-        [np.abs(g.numpy()).mean(axis=0).flatten() for g in normed_gradients]
-    )
+    gradients = [grad if grad is not None else 0 for grad in gradients]  # categorical tensors
+    gradients = [
+        np.array(_g) * _x.std(axis=0) for (_g, _x) in zip(gradients, x)
+    ]  # norm to value ranges
+    if "sample_weight" in kwargs:
+        gradients = [(_g.T * kwargs["sample_weight"]).T for _g in gradients]  # apply ev weight
+    mean_gradients = np.concatenate([np.abs(g).mean(axis=0).flatten() for g in gradients])
     return mean_gradients / mean_gradients.max()