predict_start_from_noise

[Liujingxiu23]

I do not understand this funcition in the script "diffusion_process.py". Could someone help me?

Frankly speaking，I do not fully understand the process of diffusion and the inverse process.
But for other functions, I can find corresponding equation in the paper "wavegrad" or "Denoising Diffusion Probabilistic Models". But for this function, I cannot.

And the inverse process that generate a wave seems much differ and complex than the process in
https://github.com/lmnt-com/wavegrad/blob/master/src/wavegrad/inference.py. lines 56~64.
Why is there such a difference？

[enhuiz]

Hi @Liujingxiu23 , the code seems to first calculate y_0:

WaveGrad/model/diffusion_process.py

Line 118 in d230621

y_recon = self.predict_start_from_noise(y, t, eps_recon)

and based on y_0, it calculates the mean and variance of q(y_{t-1} | y_t, y_0).

WaveGrad/model/diffusion_process.py

Line 123 in d230621

model_mean, posterior_log_variance = self.q_posterior(y_start=y_recon, y=y, t=t)

You can try to do the derivation, except for the clip_denoised, it should be identical to

https://github.com/lmnt-com/wavegrad/blob/master/src/wavegrad/inference.py. lines 56~64.

[ivanvovk]

The answer on your question has been given. If you have another questions, feel free to reopen issue or open another one.