Request: save/resume training

[rll]

Request: save/resume training

[dementrock]

Hi @123mitnik, I've pushed an experimental implementation.

For new experiment runs (must be under "stub" mode), now the snapshot files (ending in .pkl) will store a serialized version of the Algorithm object. Then, you can resume training by running

python scripts/resume_training.py PATH_TO_PKL_FILE.pkl

[dementrock]

Hi @123mitnik,

Right now the first and last are supported already. By default, if you run experiment in "stub" mode (i.e. calling stub(globals()) and then running experiment via run_experiment_lite, it should be saving pickle files under the data directory. Then, you can run python scripts/sim_policy.py PKL_FILE to replay the trained policy.

Which algorithm(s) do you need the resume training feature for? I can give a shot at implementing a primitive version.

[123mitnik]

the main focus is on Trust Region Policy Optimization and second on the list is Truncated Natural Policy Gradient
my homebrew env is doing some time consuming tasks and it will take huge amounts of time to finish the training end to end

it will be great asset if you can resume the policy training .. just a lifesaver in demanding envs

I would like to Thank You for your fast replay. And to appreciate how wonderful rllab is !

[123mitnik]

wow !!! that was FAST

Thank You so much !!! ... will start using it immediately