Thank you very much on your work on making Transformer Timeseries models easily accessible. Highly appreciated. I am currently trying to reproduce the results from the original Informer paper using the Informer Implementation from Huggingface. The authors of the original Informer paper split their datasets in 0.6/0.2/0.2 chunks. The Huggingface Informer Tutorial mentions that the val set is only “prediction length” longer than the training set and the test set only 2* “prediction length” longer. Can I simply split the input dataset like in the original paper or do I need to adjust morein the code? Currently, I am getting results MSE/MAE on the ETHh1 dataset way better than stated in the original Informer paper.
Thank you very much in advance!