KL loss scaled

[yjgeno]

Hi, just wanted to make sure whether the kl loss has to be scaled by the # of nodes, as many tutorials indicate:
model.kl_loss() / data.num_nodes

This is the source defining kl_loss:
https://github.com/pyg-team/pytorch_geometric/blob/master/torch_geometric/nn/models/autoencoder.py#L180
Is the output a mean value of kl_loss?

[rusty1s]

Yes, it has be to scaled, see here. Please also see discussions around this in tkipf/gae#20.

[yjgeno]

Thanks this is insightful.