Insights about the checkpoint_save_secs parameter

[Eriz11]

Hi all,

During my tries with rl_coach, I have used the checkpoint_save_secs parameter to save intermediate checkpoints of my training.

I have two question regarding this parameter, that I may be misunderstanding:

1. ¿What is the insight of having a time dependant saving feature? I mean, sometimes I don't really know how much time the training will take, even more if I change one of the components like the agent, environment or computer hardware. ¿Why not to have a checkpoint saving feature based on the steps taken instead on a timely basis?

2. Regarding the first question, most of the times I want to save the last state of my agent when training finishes (understood as last checkpoint). Going through the source code, I see that I could invoke the graph_manager.save_checkpoint() method to save it. ¿Is this the right path to go for saving the last state of the agent? If I don't do it, I just find time based and saved checkpoints in my directory.

Any thoughts will be very welcome.