spatial sequence prediction about patchtst

[yuqinie98]

Hi @WANGYIZI , when you say 1d space, do you mean (feature_dim * 1), that is , the temporal dimension only has one step?

[WANGYIZI]

I mean each 1D sample is divided into n segments,4 known initial features at certain time(maybe this is the one step you said) of these n segments are given.I want to know that after a long period of time,what the other 2 target features of these n segments will be.By the way,my prediction is end-to-end.It actually reflects a physical process in one dimensional space.

[WANGYIZI]

maybe my prediction is about spatio-temporal series prediction.

[yuqinie98]

Sorry I am confused. When you say 1d sample, my understanding is that either you have only one feature, so the input shape will be (1, input length), or only one time stamp with the input shape (feature_dim, 1). But since you mentioned you divided it into n segment with 4 initial feature dimension, I got a little bit confused. Could you help to clarify the shape of your raw input, input after patch, and target? Thanks

[WANGYIZI]

ok,clearly speaking,my raw input shape is(n,4),you know,n segments and 4 known initial physcial features of the n segments.The target(output) shape i hope to get is (n,2).Given 4 known initial features,I want to know, after a long period of physical process, what the other 2 target features will be.I have tried LSTM to do my task as a time series sequence prediction, but it actually is not time series sequence.So i'd like to know if your model can be competent for my task.My prediction is end_to_end.

[yuqinie98]

I see. So Transformer model is proposed to learn the attention between elements in a sequence, and in your case, it learns the relationship between all (1,4) vectors (key and query) to generate an attention map n*n. And you can then use this attention map with the value to generate an output with the shape (n,2). I would suggest you first considering the original Transformer model. Our PatchTST model is very similar just with 2 differences in patching and channel-independence.

[WANGYIZI]

oh,that is a lot.Thanks.

[yuqinie98]

No problem!