Question about using PCGrad with Uncertainity based weights as discussed in the paper for CIFAR dataset

WeiChengTseng / Pytorch-PCGrad

Pytorch reimplementation for "Gradient Surgery for Multi-Task Learning"

BSD 3-Clause "New" or "Revised" License

302 stars 42 forks source link

Question about using PCGrad with Uncertainity based weights as discussed in the paper for CIFAR dataset #11

Closed madhukiranets closed 2 years ago

madhukiranets commented 3 years ago

I was wondering how PCGrad could be used in conjunction with uncertainty-based weights. Or If I have to add scalar weights to the individual losses of each task in addition to PCGrad will the following still work,

Loss = [alpha x loss1 , beta x loss2]

where alpha and beta are weights. Will the above line add some effect due to weights added to individual losses?

lorenmt commented 2 years ago

Yes it will still work. I have tested in my latest project.