Closed sparticlesteve closed 2 years ago
I think it makes sense to make this parameter tunable. I do not see any big issues adding this to the tunable parameters list at this point in time since we are still early in the submission.
Turns out this wasn't really forgotten, but was hidden in the markdown rendering due to a bug in the table! See #495
The learning rate decay factor was mistakenly left out of the HP table for the OpenCatalyst benchmark, but is a very reasonable thing to tune and is usually allowed for other benchmarks.
I think we should allow to add a rule saying this parameter is unconstrained for HPC v2.0, despite being past the freeze deadline.