Figure out how to save and restore experiments in a cross-platform and 2-3 compatible way

TheButlah / makrl

makrl - modular algorithm kit for reinforcement learning

4 stars 1 forks source link

Figure out how to save and restore experiments in a cross-platform and 2-3 compatible way #17

Open TheButlah opened 5 years ago

TheButlah commented 5 years ago

I'm currently writing up the code to perform the cart pole q-learning experiment. I want to be able to save and restore the experiment so that it can be resumed. We should figure out how to do that.

TheButlah commented 5 years ago

Right now, I'm thinking that we can have each experiment go in a directory, and it would save everything like logs and the model checkpoints there. I'm just not sure how we would actually make it easy to restore from that - would there be a standardized process or would it be up to each model?

Additionally I'm using argparse to define all of the arguments. When restoring a model, the arguments should be loaded from that saved model, but some or all might be overwritten. How would that work?