mirror of
https://github.com/huggingface/lerobot.git
synced 2026-05-11 14:49:43 +00:00
Steven Palma
1add460678
* Fix loss normalization for padded actions in ACT, Diffusion, and MultiTaskDiT When action_is_pad masks out padded timesteps, the subsequent .mean() still divides by the total element count (including zeroed-out padding), underestimating the loss. With 60-70% padding this can cut the effective gradient signal by 2-3x. Replace mask-then-mean with mask-then-sum / valid-count for all three affected policies. TDMPC is not affected because it sums over time before averaging over batch. Fixes #3353 * linting Co-authored-by: whats2000 <60466660+whats2000@users.noreply.github.com> Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net> * Update src/lerobot/policies/diffusion/modeling_diffusion.py Co-authored-by: whats2000 <60466660+whats2000@users.noreply.github.com> Signed-off-by: Steven Palma <imstevenpmwork@ieee.org> * Update src/lerobot/policies/multi_task_dit/modeling_multi_task_dit.py Co-authored-by: whats2000 <60466660+whats2000@users.noreply.github.com> Signed-off-by: Steven Palma <imstevenpmwork@ieee.org> * Update src/lerobot/policies/multi_task_dit/modeling_multi_task_dit.py Co-authored-by: whats2000 <60466660+whats2000@users.noreply.github.com> Signed-off-by: Steven Palma <imstevenpmwork@ieee.org> * apply ACT loss normalization suggestion from review Divide by num_valid (timesteps * action_dim) instead of just timesteps, matching the diffusion/multi_task_dit fix. Addresses review from @whats2000 (https://github.com/huggingface/lerobot/pull/3377#discussion_r3106845791). * fix(test): update safetensor act --------- Signed-off-by: Maxime Ellerbach <maxime@ellerbach.net> Signed-off-by: Steven Palma <imstevenpmwork@ieee.org> Co-authored-by: Yufeng He <40085740+he-yufeng@users.noreply.github.com> Co-authored-by: Maxime Ellerbach <maxime@ellerbach.net> Co-authored-by: whats2000 <60466660+whats2000@users.noreply.github.com>