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

About conditional diffusion #64

Closed HxSun08 closed 3 weeks ago

HxSun08 commented 3 weeks ago

Thank you for adding such a great repository! It’s really helpful.

I’d like to ask if your code includes any conditional diffusion elements—for instance, generating time series data with a one-hot label. If not, could you provide some advice on where to start implementing this?

Thanks so much!

Y-debug-sys commented 3 weeks ago

Hi, all you need is Classifier Guidance or Classifier-Free Guidance . Regarding Classifier Guidance, you can train an additional classifier (x_t->label) to guide the reverse process, and the related code of Classifier-Free Guidance can be found in classifier_free_guidance.

HxSun08 commented 3 weeks ago

Thank you for your prompt reply and recommendation! I’d like to ask if your code provides an interface for the classifier. If not, could you suggest where it would be best to add the implementation? Thanks!

Y-debug-sys commented 3 weeks ago

Sorry, I don’t have any interface to provide. You might not necessarily need a classifier if you use another method (Classifier-Free Guidance). I’ll try to upload code related to classification in the future.