Early Stopping Based on Number of Repeated Best Loss

[Peyman Mohajerian]

Hi All,

I'm new to this community. I just implemented an early stopping based on number of repeated best loss values (best_loss_unchanged_threshold), e.g.:

argmin = fmin(
fn=train,
space=search_space,
algo=algo,
max_evals=16,
best_loss_unchanged_threshold=10,
trials=spark_trials)

I do realize there's 'timeout' and 'loss_threshold' early stopping strategies are already in place. Sometimes all you care about is that a certain number of times that best loss values has not changed. If there's interest also please propose a name alternative to 'best_loss_unchanged_threshold'.

Thanks,

Peyman Mohajerian