关于结果 about patchtst

[yuqinie98]

你好！或许你可以尝试减小forecasting length，然后滚动向前预测，因为看起来你的forecasting在较近的距离比较好。另外，可以尝试重点调整学习率的各个参数，这部分对TST影响较大。最后，如果你的dataset不够大，可以尝试收集更多数据或者做data augmentation。

[SimyokH]

好的好的，十分感谢您宝贵的意见！

…

---原始郵件--- 寄件人： ***@***.***> 傳送時間： 2023年5月17日(週三) 凌晨0:11 收件人： ***@***.***>; 抄送： ***@***.******@***.***>; 主旨： Re: [yuqinie98/PatchTST] 关于结果 (Issue #44) 你好！或许你可以尝试减小forecasting length，然后滚动向前预测，因为看起来你的forecasting在较近的距离比较好。另外，可以尝试重点调整学习率的各个参数，这部分对TST影响较大。最后，如果你的dataset不够大，可以尝试收集更多数据或者做data augmentation。 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[ikvision]

@SimyokH and @yuqinie98 it would be very helpful to have the discussion in English to allow other to understand the topic.
English is not my native language also, but it allows more participants into this open source community

[yuqinie98]

I agree. Thanks for addressing it @ikvision

[SimyokH]

@SimyokH and @yuqinie98 it would be very helpful to have the discussion in English to allow other to understand the topic. English is not my native language also, but it allows more participants into this open source community

Ok, thanks for your suggestion.

[1348598339]

你好！或许你可以尝试减小forecasting length，然后滚动向前预测，因为看起来你的forecasting在较近的距离比较好。另外，可以尝试重点调整学习率的各个参数，这部分对TST影响较大。最后，如果你的dataset不够大，可以尝试收集更多数据或者做data augmentation。

您好，请问滚动预测，是指训练一个网络结构，再将预测数据不断输入其中，还是滚动向前训练模型，谢谢

[yuqinie98]

Hi! It means rolling window on the data.

[CK-TD]

Hi! I'm also tring to train my own dataset with PatchTST. I'm curious about that, did the picture you upload include the train data? Or, all the timesteps(more than 1000) are the prediction horizon, and in the first 1/3 time period the model predicts with nearly 100% accuracy?