search

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

norm #7

Closed czc567 closed 8 months ago

czc567 commented 8 months ago

In your code, the "neg_one_to_one" attribute of all datasets is True. Does it mean that they are normalized to [-1, 1] during training? When calculating indicators such as Discriminative score, it is restored to [0, 1]. Is my understanding correct?

If it is correct, why do you need to do this? Will using [0, 1] during training have a big impact? Thank you for your answer！

Y-debug-sys commented 8 months ago

Hi, your right. We just continued the practices of the original DDPM paper (see its experimetal settings) and code repository. As far as I know, this trick may affect accuracy slightly, but it won't have a significant impact on the results.

czc567 commented 8 months ago

Hi, your right. We just continued the practices of the original DDPM paper (see its experimetal settings) and code repository. As far as I know, this trick may affect accuracy slightly, but it won't have a significant impact on the results.

OK, thanks