Purpose of "ALG" -> "ALGORITHM" key in config.json?

[kuttyflam]

Hi,

i was implementing some of own agents (REDQ SAC), for the LIDAR environment i was wondering what exactly is the purpose of having this:
"ALG": {
"ALGORITHM": "REDQSAC",
...

in the config? Is that just for the existing models that come with the library or can I use it for my own implementations of REDQ SAC or (other agents i may implement in the future). I saw changing it to something other than "SAC" or "REDQSAC" and running the trainer caused an error to pop up, so im not sure what it's used for and how it changes my runs, as well a how I would adapt my config when i come onto adding some other agents i wanted to test.

Thanks!

EDIT: Also might be a dumb question, but do i need to have my display on to train in the lidar based environment?

[kuttyflam]

So I see that it's used in config_objects.py and custom_checkpoints.py to update the checkpoint during training or set the parameters for the agent from the tmrldata/config, but these are directly in the library so im abit unsure how I would add my agent to it thats not REDQSAC or SAC?

[yannbouteiller]

Hi!

config.json defines the behaviour of the python -m tmrl --<option> commands, and the default arguments of the main library components.

If you implement your own pipeline (e.g., by following the long tutorial), you can ignore config.json entirely.

You can also access whatever you write in config.json through the tmrl.config module by following the same "import (...) as cfg" pattern as in config_objects.py.

[kuttyflam]

Hi!

config.json defines the behaviour of the python -m tmrl --<option> commands, and the default arguments of the main library components.

If you implement your own pipeline (e.g., by following the long tutorial), you can ignore config.json entirely.

You can also access whatever you write in config.json through the tmrl.config module by following the same "import (...) as cfg" pattern as in config_objects.py.

So i just tried to define my different config.json for now, but it seems I get an error when I run the server on the line:
from tmrl.networking import Server, Trainer, RolloutWorker

error:

pipeline.py, line 7, in <module>
    from tmrl.networking import Server, Trainer, RolloutWorker
  File ".venv\Lib\site-packages\tmrl\__init__.py", line 35, in <module>
    from tmrl.config.config_objects import CONFIG_DICT
  File ".venv\Lib\site-packages\tmrl\config\config_objects.py", line 129, in <module>
    lr_actor=ALG_CONFIG["LR_ACTOR"],
    

This is because I don't use/need a LR_ACTOR for this agent, seems like I can't seem to avoid this because its imported in the init file?

EDIT: Well it seems leaving the "SAC" section alone and adding my own stuff in addition to it fixes it since its still in the config, i guess ill use this approach for now

[yannbouteiller]

Yes, config.json is really just defining the default kwargs values used internally by tmrl. If you delete/change stuff there, things can break.

Unless you plan to add things to the tmrl codebase, you can simply import the Trainer, Server and RolloutWorker classes in your script, and instantiate them by programmatically providing the kwargs values that you want to use, as shown in the tutorial script

[kuttyflam]

Ahh thank you so much that ended up being simpler than I thought it was!