The results are different for each run

when i run svi.step()
i get the the model then predict.but the resultl is different?
i guss that maybe the seed function.
someone give me suggesstions,ths.

SVI stands for “stochastic variational inference”, and “stochastic” means that the learning process is randomized. To achieve repeatable results, you should fix that randomness via pyro.set_rng_seed(seed) for a fixed seed like 0. You’ll notice that all of Pyro’s tutorials and examples use pyro.set_rng_seed().

i get your idea.but i get the mean and std of output distribution ,these are all changed by run each run.
if i want an model to predict,the stochastic result (include mean and std )may cause useless.

If the learned mean and std are not repeatable, then it sounds like your learning process is not converging. Try decreasing step size, using a decreasing learning rate scheduler, or finding a better initialization to avoid getting stuck in local modes.