diff --git a/aimodel/src/lib/ai/components/convnext_inverse.py b/aimodel/src/lib/ai/components/convnext_inverse.py
index b866fb9..94e23c1 100644
--- a/aimodel/src/lib/ai/components/convnext_inverse.py
+++ b/aimodel/src/lib/ai/components/convnext_inverse.py
@@ -36,6 +36,7 @@ def convnext_inverse(layer_in, depths, dims):
 def block_upscale(layer_in, block_number, depth, dim):
 	layer_next = layer_in
 	
+	# Ref https://machinelearningmastery.com/upsampling-and-transpose-convolution-layers-for-generative-adversarial-networks/ to understand Conv2DTranspose
 	layer_next = tf.keras.layers.Conv2DTranspose(
 		name=f"cns.stage{block_number}.end.convtp",
 		filters=dim,
diff --git a/aimodel/src/lib/ai/model_rainfallwater_segmentation.py b/aimodel/src/lib/ai/model_rainfallwater_segmentation.py
index 7d1b2b8..8f7e16a 100644
--- a/aimodel/src/lib/ai/model_rainfallwater_segmentation.py
+++ b/aimodel/src/lib/ai/model_rainfallwater_segmentation.py
@@ -29,6 +29,7 @@ def model_rainfallwater_segmentation(metadata, feature_dim_in, shape_water_out,
 	layer_next = do_convnext_inverse(layer_next, arch_name="convnext_i_tiny")
 	
 	# TODO: An attention layer here instead of a dense layer, with a skip connection perhaps?
+	raise Exception("Error: read and implement attention from https://ieeexplore.ieee.org/document/9076883")
 	layer_next = tf.keras.layers.Dense(32)(layer_next)
 	layer_next = tf.keras.layers.Conv2D(1, kernel_size=1, activation="softmax", padding="same")(layer_next)