Question about the q_posterior function

[mikittt]

Thank you for releasing the code of this excellent work!

Regarding the below function, I couldn't figure out the usage of the q_pred function in L215 and L237.

https://github.com/cientgu/VQ-Diffusion/blob/37bbcccdd4aef1794dac645128d864a9f69ed985/image_synthesis/modeling/transformers/diffusion_transformer.py#L206

I understand q_pred as a function that takes an initial state and a time, and returns a distribution with noise at that time. However, the q_pred function in L215 receives log_x_t instead of log_x_start, while the comment says it returns q(xt|x0). In addition, I would be grateful if you would tell me which equation q_pred in L237 corresponds to.

[cientgu]

Sorry for the late reply, really busy these days.... Line215 aims to get the transition matrix. Line 233 to Line 237 is a computation trick to achieve q_posterior, briefly speaking, it first do the normalization, then leverages the q_pred function, and finally re-normalizes to get the result.

[mikittt]

Thank you for the clarification!

[cantabile-kwok]

Sorry for the late reply, really busy these days.... Line215 aims to get the transition matrix. Line 233 to Line 237 is a computation trick to achieve q_posterior, briefly speaking, it first do the normalization, then leverages the q_pred function, and finally re-normalizes to get the result.

Sorry for re-opening this, but what did you mean by "Line215 aims to get the transition matrix."? I still don't get it 😥

[fido20160817]

So, what's the specific theory behind the computation trick? It's not easy to understand the processing for me, can you give some clues behind this trick? Further, it should be log_qt = self.q_pred(log_x_start, t) in Line 125?