loss

[elilth]

Hi!
        I want to ask you, is there any difference in output loss?
return self.embedding_loss(Fx, Fe) + self.config.solver.alpha * self.output_loss(prediction, labels) + lamda * l2_norm # self.cross_loss(features, labels, keep_prob)

Is it possible here?：
self.output_loss(prediction, labels) ？

[greeness]

You should combine embedding_loss and output_loss to reproduce the loss introduced in the original paper, by considering both the embedding error and the label correlation error.

If you only use the label_correlation loss ("self.output_loss(prediction, labels), you cannot enforce the network to learn to minimize the embedding loss, which is definitely not good.