Alternate between the losses
GradNorm addresses the problem of balancing multiple losses for multi-task learning by learning adjustable weight coefficients.
- https://arxiv.org/abs/1711.02257
- https://github.com/brianlan/pytorch-grad-norm
PCGrad Gradient Surgery for Multi-Task Learning