How to use Learning Rate Annealing etc in Modulus

Physics Simulation Modulus Physics-ML Model Framework Technical Support (Modulus Only)
1

Hi,

I read in the advanced schemes:

that they are several Learning Rate Annealing mtd etc like:

Homoscedastic Task Uncertainty for Loss Weighting
SoftAdapt
ReLoBRaLo

However, there’s only an example for Neural Tangent Kernel (NTK).

So are the other mtds available in Modulus too? If so, are there examples of how to use them?

Thanks!

2

You set it in the config.yaml, e.g.


defaults :
  - modulus_default
  - arch:
      - modified_fourier
  - scheduler: tf_exponential_lr
  - optimizer: adam
  - loss: lr_annealing
  - _self_
3

Hi npstrike,

Ok thanks! I’ll try it.