Question regarding decoder inputs about patchtst

[namctin]

Hi @HasnainKhanNiazi , thanks for your interest in our paper! The concatenation is for concatenating the history with the predicted values. This is used in Informer, FEDformer and other. But with patchTST, we do not need to do so. You can safely comment out these lines when running our model. For prediction with one single target, you can also follow the script: https://github.com/yuqinie98/PatchTST/blob/main/PatchTST_self_supervised/patchtst_supervised.py where the head_type='regression'.

[HasnainKhanNiazi]

Hey, @namctin thanks for your kind reply, I just want to make sure that all the experiments that you did in the paper, those experiments were not in auto-regressive fashion where the target itself was being given in the inputs, and then while making a prediction you are just adding zeros in targets? I was exploring the code of Informer and this is what I came to know and you are also using the base implementation from Informer so please let me know if the target was also present in training or not. Thanks

[namctin]

@HasnainKhanNiazi You are correct, since we are not using the encoder+decoder architecture as in Informer but instead only considering decoder, we can't not use the same input/target structure. In our case, (input, target) is simple (x_{1:t}, x_{t+1:t+L}). Hope that is clear.

[HasnainKhanNiazi]

@namctin Yes, that makes sense. Thanks for clearing the confusion. :)