Initialize MCMC with SVI

numpyro
1

There are a couple of similar posts here about this, more so in this thread but I don’t think I have a clear understanding of the solution / proper use of init_to_value. I aim to initialize all sites in my model with the output of SVI MAP result (a dictionary which the keys are the site names and the values are the result of the optimization). From my inspection of the init_to_value function, it appears it only would work for one site? How would I initialize all sites in the NUTS kernel?

initial_values = {"site_1": site_value, ....}
kernel = NUTS(model_fn, init_strategy=init_to_value(values=initial_values))
1 Like
2

see e.g. here for an example

2 Likes
3

ah thank you!

1 Like
4

Hi,

A related question:
What is the difference between using the argument init_strategy=init_to_value(values=...) in the NUTS class and the argument init_params=... in the run method of of the MCMC class?

I.e. what is the difference between

kernel = NUTS(numpyro_model,  init_strategy=init_to_value(values=...))
mcmc = MCMC(kernel, *mcmc_args)
mcmc.run(rng_key, *numpyro_model_args)

and

kernel = NUTS(numpyro_model)
mcmc = MCMC(kernel, *mcmc_args)
mcmc.run(rng_key, *numpyro_model_args, init_params=...)

The documentation says that init_params should be in the unconstrained space, but other than that, are there any differences? I.e., are the two approaches equivalent or should I prefer one over the other, or use both?

5

I believe that is the only difference

2 Likes