Setting custom model using the new config method

RLlib
1
  • Low: It annoys or frustrates me for a moment.

Hi everyone! I’m pretty new to rllib som I’m still trying to make the config of my BC algo more efficient. I’ve been migrating from using the old-school config dict to use the new method of config=BCConfig.environment(…).rollouts(…) etc. I’ve been trying to search around for the arguments to solve the custom model import. Today I do it in the config dict like this:

“model”: {
“vf_share_layers”: True,
“custom_model”: “cfc”,
“custom_model_config”: {
“cell_size”: cell_size,
“obs_space”: obs_space,
“num_outputs”: num_outputs
},
}

Is there anyone having a link to where it says how this should be done with the new method or and example? My search so far has led me to belive it should be done in BCConfig.training(model= {dict here maybe? )

Thanks in advance.

2

Hi @R-Liebert,

Welcome to the forums.

I think it should look something like this:

model={...,
    "custom_model_config":{
        ...
     }, 
}
4

Thank you! I found the “TODO” in the algorithm_config file line 1500 that this isn’t made to an object yet.

I’m trying to convert the tensorflow example of the CfC behavior cloning (Atari Behavior Cloning - Neural Circuit Policies 0.0.1 documentation) into a following the new convention, but I’m running into all sorts of issues.