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

improve comment

Browse Source
This commit is contained in:
Michel Aractingi
2026-01-14 17:13:26 +01:00
parent 0264ac717b
commit 80b0f1aaa2
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/scripts/lerobot_train.py
+4 -2
View File
@@ -103,8 +103,10 @@ def update_policy(
# Note: Policies supporting sample weighting must implement forward(batch, reduction="none") # Note: Policies supporting sample weighting must implement forward(batch, reduction="none")
per_sample_loss, output_dict = policy.forward(batch, reduction="none") per_sample_loss, output_dict = policy.forward(batch, reduction="none")
# Apply sample weights: L_weighted = Σ(w_i * l_i) / (Σw_i + ε) # Weighted loss: each sample's contribution is scaled by its weight.
# Weights are already normalized to sum to batch_size # We divide by weight sum (not batch size) so that if some weights are zero,
# the remaining samples contribute proportionally more, preserving gradient scale.
# Weights are pre-normalized to sum to batch_size for stable training dynamics.
epsilon = 1e-6 epsilon = 1e-6
loss = (per_sample_loss * sample_weights).sum() / (sample_weights.sum() + epsilon) loss = (per_sample_loss * sample_weights).sum() / (sample_weights.sum() + epsilon)