Scheduler for alpha

[hbredin]

Would be nice to try and add a scheduler that increases alpha progressively.

This paper might be related as well:

GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks