ray tune keep_checkpoints_num is ineffective

[liususan091219]

ray tune keep_checkpoints_num is ineffective, debug to make it effective.

[liususan091219]

If use ray: override _save_checkpoint, but add with tune.checkpoint_dir(...) to make ray keep track of the checkpoint.

else: manually delete the checkpoints other than the best one in TransformersEstimator.fit(), after trainer.train().