Loss weights for non-ofg process

[busyyang]

Hi, I just find an issue about the loss weights when ofg_epoch=0:

• When do ofg, the loss_ofg and loss_reg is added with weights.

if ofg_epoch:
                '''use ofg loss'''
                ofg_lr = calc_learning_rate(epoch, max_epoch, args.ofg_lr)
                criterion_ofg = OFGLoss(iter_count=ofg_epoch, reg_weight=weights_opt[1], lr=ofg_lr)

                loss_ofg = criterion_ofg(x, y, output[1]) * weights_model[0]    # <-------- with weights
                loss_reg = criterion_reg(output[1]) * weights_model[1]          # <-------- with weights
                loss = loss_ofg + loss_reg
                loss_vals = [loss_ofg, loss_reg]

• When ofg_epoch=0, the loss_ncc and loss_reg are added without weights?

else:
                '''use ncc loss'''
                loss_ncc = criterion_ncc(output[0], y)    # <-------- without weights
                loss_reg = criterion_reg(output[1])       # <-------- without weights
                loss = loss_ncc + loss_reg
                loss_vals = [loss_ncc, loss_reg]

Is there a bug when without OFG?

[yuelinxin]

No because the default weight for unsupervised loss, i.e., NCC + regularization, is 1 : 1, so no weight is needed there. For OFG, we have OFG:Reg = 1:0.02, so there is a weight involved.