Hi, thanks for your great work! I'm currently trying to train the MK-M2Diffuser-Pick experiment using the default configuration, but encountered a shape mismatch error during training.
The error happens in the forward function of the scene_model, specifically at the line where p0 and x0 are concatenated:
x0 = p0 if self.c == 3 else torch.cat((p0, x0), 1)
However, p0 and x0 have different shapes:
p0.shape = torch.Size([294912, 3]) x0.shape = torch.Size([262144, 3])
Could you please suggest:
- Where the mismatch may originate?
- Whether a fix should be made in the data pipeline or model?
- If there's a config switch that ensures synced pos and feat sizes?
Thanks again for your work! Looking forward to your help.