Layer weights initialization

[kimzed]

I am trying to debug my model which is not working properly, and I asked myself a question regarding weight initialization.

Usually in other architecture (e.g. CNN model) we initialize the layers weight in the model class. I do not find that in any tutorial about GNNs, so I was wondering where does this take place? Am i supposed to do the initialization somewhere? I suppose it is done automatically but i would like to be sure about it.

[rusty1s]

Yes, initialization is done automatically as in regular PyTorch. You can customize them by overriding reset_parameters() or following this thread.

[kimzed]

and what is the standard initialization method? is it common practice to use the standard one?

[rusty1s]

Yes, I think it is common practice. For some operators, e.g., GCN, we make use of the initialization technique proposed by the authors.