Simple modular configuration refactoring #2
Conversation
facebook-github-bot
added
the
CLA Signed
There was a problem hiding this comment.
Thanks for submitting this!
I think it is way too complicated and convoluted at the moment. I wonder if there is a way to simplify it. Specifically, it would be nice if we still have the default config as before, but only have a folder with task specific overrides. Right now there is a lot of duplication going on and only a few parameters are changing, it would be best if we can preserve this simplicity.
README.md
Outdated
| @@ -72,7 +72,7 @@ conda activate drqv2 | |||
|
|
|||
| Train the agent: | |||
| ```sh | |||
| python train.py task=quadruped_walk | |||
| python train.py env_cfg=quadruped_walk | |||
There was a problem hiding this comment.
let's rename it to just task
cfgs/env_cfg/acrobot_swingup.yaml
Outdated
| @@ -0,0 +1,20 @@ | |||
| # task settings | |||
| task: acrobot_swingup | |||
| frame_stack: 3 | |||
There was a problem hiding this comment.
hm there is a lot of overlap going on, only a few arguments are changing. I wonder if there is a way to only have a task specific overrides and still preserve the default config.
…the relative configuration files and the parsing logic
|
As suggested, I changed the config logic to keep the default config as before and have an additional folder with task specific overrides. Also I changed the name of the folder to "task" (and the parameter specifying the task to "task_name" to avoid conflict). Let me know if this works, or if you had anything else in mind ^^ |
|
Done, added configurations for all easy and hard benchmark tasks (before I included only the medium benchmark tasks)! Let me know if any other change is required ^^ |
|
thanks, this looks good. will merge now. |
Refactored the hydra configurations to be divided into agent-specific (agent_cfg) and environment-specific (env_cfg), to ease experimentation. In particular, the only argument that needs to be overridden for any of the experiments with DrQ-v2 is env_cfg.
I have also added the relative configurations for all medium benchmark tasks, following the details from the paper.