Why the user define function of message passing in DGL (acting the same as builtin) will decrease the accuracy of node claasification

In * A Blitz Introduction to DGL

  • Write your own GNN module
    I found when use user deefine message passing method, I will lose some accuracy when training the same epoch. Can anyone help me ?

Could you provide more details such as reproducible examples for us to investigate?

Thanks for your reply! I am sorry that it’s because of the randomness of training, sorry to bother you