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?
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?
ImprintLab
