Loss function for dnorm in predcls

[jkli-aa]

[jkli-aa]

I got it. The change for $(FG+BG)/FG$ will enlarge the loss of large graph.
Sorry to bother you.

[bknyaz]

Hi, I'm glad you've resoved this issue. You are right in your last comment, but I just wanted to add that we tested the variant when the edge loss weight is fixed for all batches (equivalent to adjusting the learning rate as you mentioned). Such a variant also worked well in some cases (see the paper for details). The weight of our edge loss adapts to each batch of images/scene graphs and does not require tuning for a new dataset (where scene graphs may have different sparsity). So these properties of our loss can be helpful in practice.

[jkli-aa]

It really addressed my concerns. Thank you for your kind reply.