tsf-new-paper-taste icon indicating copy to clipboard operation
tsf-new-paper-taste copied to clipboard

A code implementation of new papers in the time series forecasting field.

Results 10 tsf-new-paper-taste issues
Sort by recently updated
recently updated
newest added

我输入的特征是B*96*4 标签是B*24*4 模型预测出来的尺度是B*16*4这个怎么处理呢?

Hi, I am one of the authors of SegRNN. Our team recently made the source code of SegRNN public: https://github.com/lss-1138/SegRNN. It can fully reproduce the results in the paper. Thank...

Thank you for providing this code! I noticed that drop last is enabled for the test set, which can lead to wrong test metrics since a few samples might get...

你好,我在使用你复现的PatchMixer跑数据集Weather预测长度为96时的mse为0.001239,远远低于论文数据,请问你在weather数据集上运行时有无遇到这个问题。

self.lucky = nn.Embedding,这个lucky层去掉会导致性能下降?好像并没有啊,后面都没用上。还有就是位置编码和通道编码是靠torch.randn就实现了? (俺是初学者,如果问题太蠢勿喷啊)

你好,PatchMixer在复现ETTh1的结果上与原论文存在差距,括号内为复现结果,请问一下你是否存在相同问题? ![1701757781492](https://github.com/hughxx/tsf-new-paper-taste/assets/64340331/a08500aa-e260-4e89-973c-a74b60aa2d34)

how to get prediction for real goal number rather than a matrix; I want to know the meaning of the final generated image. Can you tell me? Thank you very...

Have you noticed how large batch he set in the paper?

我在阅读代码的时候,发现在数据输入之前有一次scale,而输出的时候并没有反标准化,这使得mae和mse都非常小,我又去计算了mape,误差很大,原论文中并没有提到mape,这是否说明模型本身并不好,只是标准化这一操作使mae这一类非百分比误差看起来很小而已,期待您的回复和解答 >>>>>>>testing : loss_flag2_lr0.0001_dm512_test_Informer_ETTh1_ftM_sl96_pl96_p16s8_random2021_0