During the inference phase ,there were copy mismatches in the CFG and errors in the noise sampling sapce.

[Jerry-Auto]

1. The dataset uses repeat_interleave,but flag and x are directly repeated,resulting in inconsisitent coping.for example,[1,2,3,4] to[1,1,2,2,3,3,4,4] and[1,2,3,4,1,2,3,4],mismatched.
2. the initial noise sampling space must be consistent with that used during training;that is ,sampling must be performed before segmentation,and then segmentation must occur, rather than sampling directly,in the segmented sapce.
3. x_init = torch.randn((B, self.action_num, self.planner_params['action_len'], self.planner_params['state_dim']), device=self.device)
   to x_0 = torch.randn((B,1,self.planner_params['future_len'], self.planner_params['state_dim']), device=self.device)
   x_init = traj_chunking(x_0, self.planner_params['action_len'], self.planner_params['action_overlap'])
   x_init = torch.cat(x_init, dim=1)

[hanzh0816]

@Jerry-Auto Hi, I noticed the same issue with repeat_interleave as well. It seems to affect the training process (specifically, the batched data), whereas it does not occur during inference. Have you tested training the model with the cond/uncond data generated by repeat_interleave strictly following the original repository's code? If so, how did it perform, and did it achieve the performance gains claimed in the paper?

[roginsteam]

@hanzh0816 I think it doesn't affect the training process, because the 'data.repeat(2)' only occurs in the inference process.

[hanzh0816]

@roginsteam thanks for the replay. It seems this issue won't have much impact, since inference only takes one sample at a time as input. I'm not sure if my understanding is correct.

[Jerry-Auto]

@Jerry-Auto Hi, I noticed the same issue with repeat_interleave as well. It seems to affect the training process (specifically, the batched data), whereas it does not occur during inference. Have you tested training the model with the cond/uncond data generated by repeat_interleave strictly following the original repository's code? If so, how did it perform, and did it achieve the performance gains claimed in the paper?

Replication mismatch only affects the CFG during inference, but not during training. If you set the weight of CFG to 1, meaning it will not be used, it will still work normally.Besides, the initial noise sampling space during inference must be corrected first, otherwise normal trajectories cannot be generated.