关于mask prompt的在3D数据集上/视频序列的训练方法Training Methods for Mask Prompting on 3D Datasets / Video Sequences

[sove45]

我注意到在您的代码中也存在着
I noticed that your code also contains

在训练3d数据集的相关代码，可是当我使用这个函数来尝试利用第一帧的mask作为prompt来进行训练的时候，发现出现第一帧没有梯度的报错，我通过debug发现是在代码中，对于有mask prompt的帧直接输出输入mask导致的，
In the code related to training on 3D datasets, when I tried to use the mask from the first frame as a prompt for training, I encountered an error indicating that the first frame lacked gradients. Through debugging, I discovered that the issue was caused by the code directly outputting the input mask for frames with mask prompts.

所以我再训练的过程中不让输入mask作为pred直接输出，而是让这个语句判断为false，走下面的分支
Therefore, during the training process, I modified the code to prevent the input mask from being directly used as the prediction output. Instead, I ensured that the condition evaluates to false so that the execution follows the alternative branch.

可是训练结果不收敛，我注意到您在论文中有提到mask的训练方法，请问应该怎么在3d数据集或者视频中利用mask prompt中来进行训练呢
However, the training results are not converging. I noticed that in your paper, you mentioned a method for training with masks. Could you advise on how to effectively use mask prompts during training on 3D datasets or video sequences?

[sove45]

请问我的训练方法在哪里存在着问题呢，或者是计算prompt loss的方法存在问题
Could you help identify where the problem might lie in my training method, or is there an issue with how I compute the prompt loss?