Non-Tune Early Stopping criterion

[jmsckv]

Functionality that stops a trial if validation performance has not improved after X epochs.
Should avoid blocking of resources (especially at the beginning of training).

Tune has sth. like a Stopper class that should be used.
Implement as a configurable hyperparameter, default behaviour could be after a fraction of max epochs.

Sth like:
s = Stopper()
grace_period_no_improvement = config.get('early_stop_no_improvement', int(config['max_t']/10))

and then in every iteration:
if updated_best:
cnt = 0
else:
cnt += 1
if cnt > grace_period_no_improvement:
s.stop_trainable()