Some datasets missing one time stamp compared to the UCR

[hadifawaz1999]

Hello,

First thank you very much for the simplicity of reproducing your work, the notebook is very straightforward, and I really enjoyed reading your work.

While reproducing the generation for all the datasets using the notebook and the pre-trained models you provide, I noticed that for 33 datasets the generations have 1 time stamp less than the original dataset. The generations are correct though I plotted them and they have the right form but when i print the shape they have one time stamp less. Here is the list of the datasets:

['ArrowHead',
'Car',
'CinCECGTorso',
'DiatomSizeReduction',
'EthanolLevel',
'FaceAll',
'FacesUCR',
'Fish',
'FreezerRegularTrain',
'FreezerSmallTrain',
'Fungi',
'GesturePebbleZ1',
'GesturePebbleZ2',
'Ham',
'HandOutlines',
'InsectEPGRegularTrain',
'InsectEPGSmallTrain',
'Lightning2',
'Lightning7',
'MedicalImages',
'OSULeaf',
'PickupGestureWiimoteZ',
'ShakeGestureWiimoteZ',
'SmoothSubspace',
'SonyAIBORobotSurface2',
'Strawberry',
'ToeSegmentation1',
'ToeSegmentation2',
'Trace',
'UWaveGestureLibraryAll',
'UWaveGestureLibraryX',
'UWaveGestureLibraryY',
'UWaveGestureLibraryZ']

Can you let me know if:

1- This is my wrong doing or do you have the same thing when producing them yourselves.
2- If you do notice the same thing, do you know the reason your model does that ?

Thank you for your response in advance.

[danelee2601]

Hi! Thanks for your interest. The one timestep missing problem occurs due to the STFT (short-time Fourier transform) process.
Let's say we have time series x with a length of 101, then after it undergoes STFT and ISTFT, its length would become 100 instead of 101 (maybe not exactly 101 -> 100, but the problem is certainly due to the STFT operation).