admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-07-25 18:56:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add Important sampling, only use replacement, remove beta smoothing

Browse Source
This commit is contained in:
Pepijn
2025-03-14 13:09:05 +01:00
parent 6a8be97bb5
commit 17d12db7c4
3 changed files with 55 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
lerobot/common/policies/act/modeling_act.py
+2 -3
View File
@@ -161,7 +161,6 @@ class ACTPolicy(PreTrainedPolicy):
l1_loss = elementwise_l1.mean()
# mean over time+action_dim => per-sample array of shape (B,)
l1_per_sample = elementwise_l1.mean(dim=(1, 2))
if self.config.use_vae:
@@ -175,13 +174,13 @@ class ACTPolicy(PreTrainedPolicy):
loss_dict = {
"l1_loss": l1_loss.item(),
"kld_loss": mean_kld.item(),
"per_sample_l1": l1_per_sample, # shape (B,)
"per_sample_l1": l1_per_sample,
}
loss = l1_loss + mean_kld * self.config.kl_weight
else:
loss_dict = {
"l1_loss": l1_loss.item(),
"per_sample_l1": l1_per_sample, # shape (B,)
"per_sample_l1": l1_per_sample,
}
loss = l1_loss