Predicting node degree

[WilliamKRobert]

Hi, I am making a simple experiment in which GCN is used to predict node degree (it sounds silly to do that as node degree is easy to compute; it is just for understanding the behaviour of GCN anyway). The only change to original code is replacing loss function with L1 loss and modify GCN to output regression predictions instead of classification probability. However, the accuracy of prediction is bad; the network gives completely wrong predictions. I would expect that this task was super simple for GCN to learn. Any suggestions?