Saving model / policies / weights after PPO training with a custom TFModelV2

RLlib Checkpointing, Restoring
1

Hi Ray community,

After training a PPO agent, I would like to export model / weights / policy to be able to apply the learnt policy on other environment instance.

What is best way to do so? Which aspects do need to be considered when using a custom TFModelV2?

I am asking that because I have experienced in the past that there are many “hidden” relations and dependencies when working with rllib. For example, I have already observed an open issue in rllib which makes me assume that keras may also play a role in that game …

My request is not related to this thread, as in that case the export / saving works fine.

Thanks in advance for your advice!

2

Is there any update on the topic?

3

Any update on that topic?

4

I’m not sure if this link will be helpful to you
It‘s about transfer learning