作者你好，请问算子到底是啥，是如何得到的呢？通过查看源码发现时不变部分的算子是可学习参数的SVD的左右奇异矩阵相乘，时变部分的算子是可学习参数的最小二乘法的解，请问为什么是将这两个当成算子？

您好，我对于时间序列非平稳带来的问题一直存在一些疑惑，时间序列非平稳是指序列的统计信息随时间变化，可是这到底是如何影响神经网络模型的预测呢？神经网络预测输入的是历史序列，输出的是未来序列，只要这个映射关系是存在一定规律的，那么神经网络应该就能够在一定程度上学到这种关系，序列的非平稳是如何影响这种映射关系的呢？为什么平稳的序列就一定更好呢？（白噪声也是一种平稳序列，可是却完全无法预测。

As I tried Koopman operator in financial data, it did not show very good property. I am curious about reproduction in other dataset.

Hello,

Thank you for this project and contribution.

Currently, the model predicts only one batch given a CSV file. Is there a convinient way to modify it so that it can predict on the entire dataset, similar to how predictions are made on the test set?

Dear author,
In your code, I found a mistake where the code does not match the paper, like picture

Therefore, after I fix this bug, my mse and mae decline a little

Looking forward to your reply

利用koopman理论有限维逼近无限维，可以处理高维度的时间序列，看到作者提出的算子适应适合长期预测，不知道是否适用于短期预测呢？

RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.请问这是怎么回事，我在linux里面就可以正常执行代码

RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.

when I run the run.py, Program encountered operational issues,so how to resolve it?

Thanks for your amazing work!
I don't quite understand why you set pred_len=1 in
https://github.com/thuml/Koopa/blob/main/models/Koopa.py#L92.

I am running experiments on "ETTh2_96_48".

torch版本：1.10.0+cu113
报错信息：NotImplementedError: the derivative for 'linalg_lstsq' is not implemented.
出错位置：

And did you test/comfirm the multi-covariable features can help to improve the predict accuracy?

Thanks!

I'm confused about the code in the model part, that is, line 74 - line 98 in models/Koopa.py.

Why isn't the variable z_pred get from one_step_forward() function put into the list z_preds?

There is a time interval between variable x and variable y, and the least square method is used to calculate $K_{\mathrm{var}}$ . Next, the value z_pred is calculated with the last segment of the look-back window and $K_{\mathrm{var}}$ as input, corresponding to #L85. This z_pred variable seems not belong to the look-back window, but the predicting window. So why isn't this z_pred variable in the z_preds ?

I try to change the code snippets of line 94 - line 95 to this:

        z_preds = []
        z_preds.append(z_pred)
        for i in range(1, pred_len):

And the result gets worse, well, that makes me confused.

作者您好，这篇文章提到了动力学系统，时间序列的不平稳和非线性研究方向贵课题组后续还会有涉及吗？

Hello Author，I am curious about two component features after Fourier Filter, can you tell me the details or show me the code how you did the visualization study on features and how can I calculate Degree of Variation

作者您好，我阅读代码的时候看到 dataloader里已经使用StandardScaler()做了处理，为什么在Model的forward里面还要做Stationarization呢？

hello,when I run your code,its has a error, when runnging at "loss.backward()", its has a mistake :"RuntimeError: the derivative for 'linalg_lstsq' is not implemented." How to resolve it,can you help me? what's the pytorch version?