Tutorial0 Unnormalize and Question on FFT term in loss funciton

Y-debug-sys / Diffusion-TS

[ICLR 2024] Official Implementation of "Diffusion-TS: Interpretable Diffusion for General Time Series Generation"

MIT License

186 stars 27 forks source link

Tutorial0 Unnormalize and Question on FFT term in loss funciton #3

Closed littlestone111 closed 8 months ago

littlestone111 commented 8 months ago

The commented code seems to work for datasets other than sine. ori_data = np.load(os.path.join(dataset.dir, f"{dataset_name}_normtruth{seq_length}_train.npy"))

Maybe there is something with unnormalize_to_zero_to_one?

Want to confirm the FFT loss term is utilized for each feature (time series) not each time step.

Y-debug-sys commented 8 months ago

For the first question, it has been unnormalized to [0, 1] before saving: Screen And regarding the second question, see the following screenshot： Simulate The output is symmetrical in the second dimension (with length 24). Hope this answer can solve your problem.

littlestone111 commented 8 months ago

Thank you for the fast reply.

Could you please elaborate a little bit on each dimension of ts_simulate. Is it (Batch, sequence length, # of variable)?

If yes, I guess it is FFT on each variable and on its time domain.

Y-debug-sys commented 8 months ago

Yes, the dimension is the same as the input tensor.

littlestone111 commented 8 months ago

Cool. Thank you for your time.