Loss Function Problem

[lazyeden]

Thank you for your work. I have the following questions to discuss with you:

1. why does the loss mention in equation 2 of the paper need to sum the t=2...T losses, while the code implementation only samples one time step t at each batch: t, weights = self.schedule_sampler.sample(micro.shape[0], dist_util.dev())
2. why is the MSE loss written in equation 2 of the paper ||EMB(w)-f(z1,1)||, while the code implements ||EMB(w)-f(z0,0)||:
   t0_mask = (t == 0)
t0_loss = mean_flat((x_start_mean - model_out_x_start) ** 2)
terms["mse"] = th.where(t0_mask, t0_loss, terms["mse"])
3. what is the meaning of predict_xstart, and how is the result of model_out_x_start different when it is True or False
4. Is the rounding process reversible? In other words, can I generate text and then work backwards to get the embedding before rounding, or even go forward to get the noise of each step of the denoising process?