Unconditional SSSD model.
lqh52 opened this issue · 3 comments
Thanks for the interesting work and for releasing your code!
Can you also release the code of that you implement the unconditional training, following the procedure proposed in RePaint paper, please!
Many thanks!
Hello @lqh52,
Altough is a bit computational more expensive, the simplest way to achieve the unconditional training and the way it was done is to blank out completely the conditional signal that is passed to the model, thus you will compare MSE in regards a true signal and a generated signal that was made out of only noise.
Hope this helps
I found this project from googling around the topic of generating time series data using diffusion models. I also need to do the unconditional kind and your project looks like a good place to start. By blanking out do you mean set the values to zero? thx
exactly, if you want to avoid adjusting too much the code, (deletin the mask and and adjusting shapes into the model pipeline), the fastest way is to set them to zero