ASHAScheduler already early stops underporfoming trials. This means that badly performing models wil be cut shoprtly. However overall it will still tune up to max_t even when no trial is improving. So to early stop the overall run maybe the RunConfig stop criteria (it can be a function) can be explored.
mathysgrapotte
commented
1 month ago
ASHAScheduler already early stops underporfoming trials. This means that badly performing models wil be cut shoprtly. However overall it will still tune up to max_t even when no trial is improving. So to early stop the overall run maybe the RunConfig stop criteria (it can be a function) can be explored.