search

huggingface / lerobot

🤗 LeRobot: Making AI for Robotics more accessible with end-to-end learning
Apache License 2.0
7.55k stars 711 forks source link

General support for LR scheduler and optimizer #104

Open qgallouedec opened 6 months ago

qgallouedec commented 6 months ago 
          Okay, I see why you need these "hacks". Ideally what we want here is:
  1. General support for LR scheduling. Therefore the lr scheduler params should move out of the policy config.
  2. Policies have an optional get_optimizer_param_dicts. Might need more thinking through for the general design. Maybe we can keep these ideas out of this PR though. But what are your thoughts?

_Originally posted by @alexander-soare in https://github.com/huggingface/lerobot/pull/99#discussion_r1579644384_

alexander-soare commented 6 months ago

Note that this was partially addressed by https://github.com/huggingface/lerobot/pull/121 where the LR scheduling was moved out of the policy config.

All that's left for this issue is to think about how to have the optimizer parameter selection not hard-coded.