微调之后的模型，用来预测HOMO能级，在边界出现预测值不变的情况

[jianing1997]

Summary

模型在预测分子HOMO能级时，当能级低于-7.5 eV后，出现预测瓶颈，即无论真实值怎么变化，预测值都是-7.5eV左右。这是什么原因?

Uni-Mol Version

Uni-Mol Tools

Details

模型在预测分子HOMO能级时，当能级低于-7.5 eV后，出现预测瓶颈，即无论真实值怎么变化，预测值都是-7.5eV左右(图1和图2）。这是什么原因?
图三是训练数据集

[Naplessss]

If possible, you can attach your data so that we can a quick check and reproduce, OR email me [email protected]

[ganfisher]

I encountered the same problem.

[Naplessss]

Hi ganfisher, you can share the dataset if possible, we will check and reproduce this. BTW, i think it is due to scalar settings.

[ganfisher]

Hi ganfisher, you can share the dataset if possible, we will check and reproduce this. BTW, i think it is due to scalar settings.

[email protected]

Thank you, and I have sent you an e-mail.

[emotionor]

@ganfisher @jianing1997 Hi, the issue has been fixed in #313. During the training process, the variance of the output changes after applying dropout. This change in variance can lead to a shift in the output after being mapped through a non-linear layer. Meanwhile, the output of a regression problem is an absolute value, which makes it highly sensitive to such changes. By using a linear layer and removing dropout, this issue can be resolved. However, removing dropout might also introduce other problems, such as overfitting. Therefore, if any issues arise, we welcome discussions and solutions.