Loss of Convergence in Iterative Loop (iter.00) Despite Initial Convergence Rates

[Shangguanying992]

While using DeepKS, I observed an anomaly regarding convergence rates during the initial iteration (iter. init) and subsequent iterative loop (iter.00).

During the first iteration (iter. init), the convergence rates were satisfactory with a training set rate of 0.77 and a testing set rate of 0.88.
However, upon entering the iterative loop (iter.00), the convergence rates dropped to 0, rendering further computation impossible.
Another user reported this issue as well.

Additional Context:
Software: DeepKS, abacus 3.5.3
The DeePKS water_single example can run correctly on my machine.

Thank you so much for your help in addressing this issue.

[Shangguanying992]

example.zip

[y1xiaoc]

Thanks for the question. This happens occationally when the first iteration is trained so "hard" that it overfits the data. I would suggest reduce the number of training epoch for first iteration.

[Shangguanying992]

I tried changing the value of n_epoch from 500 to 10, 100, but it seems that the situation does not change. Are there any other solutions?

[y1xiaoc]

To clarify, you may want to reduce the training length for the init train (iter.init), which is controlled by paramteres in init_train, like here in the example. I think the one you are changing is for the following iterations.

[Shangguanying992]

I changed it in params.yaml, like the file below. Do I truly understand? Or do I need to add a file named args.yaml?
params.yaml.zip

[y1xiaoc]

You may need to use the args.yaml. You can check the iter.init folder to see if the training has been redone with reduced epochs.

[Shangguanying992]

I‘m sure that the training has been redone with a reduced epoch, but the convergence rate is still 0. In this case, is there a recommended n_epoch value?

[ouqi0711]

You may try symmetrizing the descriptors via modifying the init_train block in params.yaml as suggested here. (See last part of params.yaml)

[Shangguanying992]

I have tried to modify the params.yaml as suggested, but it still has the same problem. I changed n_epoch and start_lr.

[xuwenli-nku]

在使用 DeepKS 时，我观察到初始迭代 （iter. init） 和随后的迭代循环 （iter.00） 期间收敛率出现异常。

在第一次迭代 （iter. init） 期间，收敛率令人满意，训练集率为 0.77，测试集率为 0.88。然而，在进入迭代循环 （iter.00） 时，收敛率下降到 0，无法进一步计算。另一位用户也报告了此问题。

附加上下文： 软件：DeepKS、abacus 3.5.3 DeePKS water_single示例可以在我的机器上正常运行。

非常感谢您帮助解决此问题。

您好，我是一名刚开始学习deepks的初学者，想向您请教如何才能复现官网中给出的water_single和water_cluster的例子呢，我在复现的时候err文件中给出了pyscf版本与目前deepks库不匹配的问题，还有就是这个deepks的任务可以通过提交到远程服务器slurm上跑通嘛，因为我被困扰了很久，实在找不到解决方法，希望您能给我一点建议，麻烦了！