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
Thank you very much for your update and brings us new things!
The performance after the update has indeed improved, but there are still some problems. I have some questions about the new action heads:
1.Problem of DiffusionActionHead
The updated DiffusionActionHead seems to be more unstable than before. After 50,000 steps of fine-tuning in the ALOHA environment, it still cannot be simulated normally, and an error of the MuJoCo will appear.
2.Problem of UNetDDPMActionHead
Thank you for providing the new UNetDDPMActionHead! But there seems to be some problems: In the ConditionalUnet1D of unet.py, the Downsample1d and Upsample1d will cause changes in dimensions, especially when the dimensions are odd, which may lead to wrong results. For example, when the dimension of action is (1, 1, 25, 512), Downsample1d will change it to (1, 1, 13, 1024), but in the Upsample1d, it will become (1, 1, 26, 512), which cannot match the hidden_reps dimension (1, 1, 25, 512). Changing action_horizon to 64 can solve this problem, but it cannot inference correctly. The following are my settings when finetuning:
Thank you very much for your update and brings us new things!
The performance after the update has indeed improved, but there are still some problems. I have some questions about the new action heads:
1.Problem of
DiffusionActionHead
The updated
DiffusionActionHead
seems to be more unstable than before. After 50,000 steps of fine-tuning in the ALOHA environment, it still cannot be simulated normally, and an error of the MuJoCo will appear.2.Problem of
UNetDDPMActionHead
Thank you for providing the new UNetDDPMActionHead! But there seems to be some problems: In the ConditionalUnet1D of
unet.py
, theDownsample1d
andUpsample1d
will cause changes in dimensions, especially when the dimensions are odd, which may lead to wrong results. For example, when the dimension ofaction
is(1, 1, 25, 512)
,Downsample1d
will change it to(1, 1, 13, 1024)
, but in theUpsample1d
, it will become(1, 1, 26, 512)
, which cannot match thehidden_reps
dimension(1, 1, 25, 512)
. Changingaction_horizon
to 64 can solve this problem, but it cannot inference correctly. The following are my settings when finetuning:The text was updated successfully, but these errors were encountered: