search

Y-debug-sys / Diffusion-TS

[ICLR 2024] Official Implementation of "Diffusion-TS: Interpretable Diffusion for General Time Series Generation"
MIT License
171 stars 24 forks source link

Cross correlation Loss #37

Closed ilijanikolov closed 3 months ago

ilijanikolov commented 3 months ago

Hello,

For the cross correlation loss, I was wondering why there was an extra factor of 10 in this function:

def compute(self, x_fake): cross_correl_fake = cacf_torch(self.transform(x_fake), 1).mean(0)[0] loss = self.norm_foo(cross_correl_fake - self.cross_correl_real.to(x_fake.device)) return loss / 10.

Is this 10 hard coded? Should it be the number of features?

Thank you!

Y-debug-sys commented 3 months ago

Hi, this 10 is just used for scaling down the main results of our paper, and it has no other special meaning. The metirc's detailed description can be found in the paper [Conditional-Sig-Wasserstein-GANs], and we reuse its official code. Thanks!