I happened to found when I use following codes to produce new samples of docs, I got the same doc samples every time.
predictive = pyro.infer.Predictive(self.model, guide=self.auto_guide, num_samples=num_samples)
svi_samples = predictive(docs=None, labels=labels, args=args)
return svi_samples
Then I refer to the AmotizedLDA example, the same thing happen untill I change the number in set_rng_seed(), such as from set_rng_seed(0) to set_rng_seed(1).
Is this really normal?