admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-06-18 16:57:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adapt state/action projections instead of full-finetuning

Browse Source 
There doesn't seem to be a benefit to fully fine-tune these layers
over just adapting them, so we do that instead.
This commit is contained in:
nemo
2025-12-16 18:29:21 +01:00
parent 51876e7f55
commit 3bc557552b
1 changed files with 7 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/scripts/lerobot_train.py
+7 -23
View File
@@ -125,35 +125,19 @@ def update_policy(
def get_default_peft_configuration(policy_type):
"""Build a PEFT configuration for the given policy type assuming that we train a policy from scratch
(i.e. only parts of it are pre-trained) and not from a checkpoint. This means that some layers are targeted for
full fine-tuning via `modules_to_save`, e.g. `state_proj` for SmolVLA which would otherwise be randomly initialized.
"""Build a basic PEFT configuration for the given policy type assuming that we train a policy from a checkpoint."""
common_projections = "state_proj|action_in_proj|action_out_proj|action_time_mlp_in|action_time_mlp_out"
Users can still override the full fine-tuning of these layers by passing `--peft.full_training_modules=[]`.
"""
if policy_type == "smolvla":
return {
"target_modules": r"(model\.vlm_with_expert\.lm_expert\..*\.(q_proj|v_proj))",
"modules_to_save": [
# these are initialized randomly and need full-finetuning
"state_proj",
"action_in_proj",
"action_out_proj",
"action_time_mlp_in",
"action_time_mlp_out",
],
"target_modules": rf"(model\.vlm_with_expert\.lm_expert\..*\.(q|v)_proj|model\.({common_projections}))",
"modules_to_save": [],
}
elif policy_type in ("pi0", "pi05"):
return {
"target_modules": r".*\.gemma_expert\..*\.self_attn.(q_proj|v_proj)",
"modules_to_save": [
# these are initialized randomly and need full-finetuning
"state_proj",
"action_in_proj",
"action_out_proj",
"action_time_mlp_in",
"action_time_mlp_out",
],
"target_modules": rf"(.*\.gemma_expert\..*\.self_attn.(q|v)_proj|model\.({common_projections}))",
"modules_to_save": [],
}
return {"modules_to_save": None}