cascade.core.DiscoverScheduler.on_validation_end

DiscoverScheduler.on_validation_end(trainer, pl_module)

Main logic of the scheduler

Return type:

None

Note

The scheduler will adjust the hyperparameters of the model according to the gradient norms of the likelihood, sparse, acyclic, and kernel gradients, each time the early stopping criteria is met. The adjustment is based on the ratio of the likelihood gradient norm to the other gradient norms. The scheduler continues until the constraint is satisfied or constraint violation stops improving for a consecutive patience times.