Problem of parameters with Wifi&FMCW

[whitecloudy]

Hello,

I have recently attempted to perform inference for Wi-Fi and FMCW using the provided code. However, I have noticed an issue with the parameters used in these experiments.

According to the paper, the value of self.alpha_bar[T], corresponding to the gamma_hat(T) in the T-step, should be converges to 0. However, the noise_schedule for both Wi-Fi and FMCW seems to be set to values that are too small, causing self.alpha_bar[T] to become significantly larger than it should be.

As a result, when running inference with "fast_sampling"(which starts from gaussian noise) instead of "native_sampling", the results of the paper cannot be reproduced. This outcome is fundamentally incorrect, based on both the basic principles of Diffusion models and the claims of your paper.

Could you please review this issue and provide clarification? I look forward to your response.

Thank you!
@Guoxuan-Chi