I am training an PINN where I have initial, final, and transient conditions (turned into losses). I am training the model with a net loss = loss_initial + loss_final + loss_transient. How do I give priority to a loss for backpropagation? For example, scaling higher the contribution of one loss in the gradient with respect to the weights. Scaling the individual loss seems to have negligible effects.
Scaling an individual loss by some factor, e.g. loss = a * loss1 + b * loss2 should scale the gradient backpropagated for that branch proportionally.