Checkpointing Features

[bbalaji-ucsd]

Right now Coach saves checkpoints every X seconds. It would be great if I can save checkpoints every X iterations or save checkpoints if it reaches an evaluation reward threshold.

It will also help if I can save checkpoints as a protobuf file, so I can use the saved model independent of Coach.

[safrooze]

So you want to save a checkpoint at a certain reward threshold but continue training? Is this in case you notice overfit later on and can go back to an earlier set of parameters?

[bbalaji-ucsd]

Yes, to both questions. Although flexibility to stop training after a reward threshold also helps. In the latter case, it would help to save a checkpoint at the end of training.

[safrooze]

There already is ability to stop training when a specific reward is achieved: "-asc" or "--apply_stop_condition" will stop early if a specific reward is hit and saves a checkpoint at the end.

[bbalaji-ucsd]

Ok, that solves one of the problems then :)