Hi, thanks for the work. By the code, I was wondering why it argsorted the pre-generated trajectories given the GT trajectory and set the closest one as the first, which is then used directly as the data input? Futhermore, when calculating the loss, the labels are all zeros, which is the first, closest one in the trajectories. Isn't it a hack? The model just needs to know the score of the first trajectory is always the highest. Thanks for the explanation in advance.
|
first_stage_trajs = first_stage_trajs[np.argsort(expert_error_1)] |
|
label = torch.zeros(scores.shape[0], dtype=torch.long).to(scores.device) |
Hi, thanks for the work. By the code, I was wondering why it argsorted the pre-generated trajectories given the GT trajectory and set the closest one as the first, which is then used directly as the data input? Futhermore, when calculating the loss, the labels are all zeros, which is the first, closest one in the trajectories. Isn't it a hack? The model just needs to know the score of the first trajectory is always the highest. Thanks for the explanation in advance.
DTPP/data_process.py
Line 252 in 4d721b6
DTPP/train_utils.py
Line 73 in 4d721b6