I am interested in the objective function (equation (4) in the paper).
The last term is to maximize diffusion when data is sampled from OOD data, but in the code, both criterion2 used for loss_in, and loss_out are the same. loss_in = criterion2(predict_in, label), loss_out = criterion2(predict_out, label).
As I understand, both losses are used to minimize net output vs label, I can't see the max term.
Hi Kong,
Thank for your sharing code!
I am interested in the objective function (equation (4) in the paper). The last term is to maximize diffusion when data is sampled from OOD data, but in the code, both criterion2 used for loss_in, and loss_out are the same.
loss_in = criterion2(predict_in, label),loss_out = criterion2(predict_out, label).As I understand, both losses are used to minimize net output vs label, I can't see the max term.