You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.
You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Released LoRA checkpoints lack the per-dataset displacement buffers needed by decode_positions #1
Hi, thanks for releasing FloMo and the checkpoints.
I tried to decode motion-token predictions to 3D positions with the released LoRA checkpoints from shitijgovil/FloMo and could not, because the per-dataset displacement scale buffers are not in the files.
What I see
ma/flomo_ma.pt holds 614 tensors under model and vma/flomo_vma.pt holds 616; neither contains per_ds_cumulative_disp_pct_low / per_ds_cumulative_disp_pct_high (nor the other per_ds_* buffers registered in flomo/models/modalities.py:449-456).
So after loading, those buffers keep their empty initial shape (0, 3, 1, 1, 1), and MotionTokensWanPretrainedModality.decode_positions (modalities.py:691) fails at the index_select on line 710.
The top-level normalizer_stats entry holds the action normalizer (molmoact2, realyam: p02/p98 of eepose_6drot), not these track displacement percentiles.
Likely cause
save_checkpoint in flomo/utils/train_utils.py:298-300 keeps, when lora_enabled, only state-dict keys that start with a requires_grad parameter name. Buffers are never parameters, so they are dropped from every LoRA checkpoint.
Ask
Could you publish the track_stats used for training (cumulative displacement low/high percentiles per dataset: egoverse, molmoact2, realyam, realyam_human), or checkpoints saved with these buffers? For future checkpoints, adding model.named_buffers() to the LoRA branch of save_checkpoint would keep them.
Activity