Hi! @yilundu @vacancy
Thank you for releasing the code. After examining and running it, I have several questions regarding the prediction algorithm implementation:
-
In the paper, solution sampling appears to be performed via energy gradient descent. Does this correspond to the self.opt_step function within the p_sample_loop method in class GaussianDiffusion1D?
-
If so, I noticed that an additional p_sample operation is executed before the gradient descent process. This step doesn't seem to be described in the original sampling algorithm presented in the paper. Could you explain the purpose of this additional operation?
It would be extremely helpful if you could provide detailed explanations or references to the paper of the sampling process workflow and design choices in the code implementation. Thank you!
Hi! @yilundu @vacancy
Thank you for releasing the code. After examining and running it, I have several questions regarding the prediction algorithm implementation:
In the paper, solution sampling appears to be performed via energy gradient descent. Does this correspond to the
self.opt_stepfunction within thep_sample_loopmethod inclass GaussianDiffusion1D?If so, I noticed that an additional
p_sampleoperation is executed before the gradient descent process. This step doesn't seem to be described in the original sampling algorithm presented in the paper. Could you explain the purpose of this additional operation?It would be extremely helpful if you could provide detailed explanations or references to the paper of the sampling process workflow and design choices in the code implementation. Thank you!