Question about q_posterior.

[guyuchao]

Although you have noticed us that we can use x_t instead of x_tmin1, I have trouble understanding why it can hold.
# unnormed_logprobs = log_EV_qxtmin_x0 + log q_pred_one_timestep(x_t, t)
# Note: _NOT_ x_tmin1, which is how the formula is typically used!!!

Instead, I want to to follow the original formula that first sample x_tmin1 based on x0, then use this to compute q(xt|x_{t-1}). Is the following code correct? And have you tried experiment with the original formula?
log_x_tmin = self.log_sample_categorical(log_EV_qxtmin_x0)
unnormed_logprobs = log_EV_qxtmin1_x0 + self.q_pred_one_timestep(log_x_tmin, t)

[ehoogeboom]

The comment in the code there is a bit confusing, I'll remove that.

Your last suggestion though is to sample x_tmin1, that is doing something different. In the original formula's we don't want to sample, we want to compute the posterior closed-form.

[guyuchao]

Great thanks.