Gibbs Sampling

[luiarthur]

If providing full conditionals, this works.

def model(ctx: Context, y, analytic=False):
  mu_prior = Normal()
  if analytic:
    mu = ctx.cached("mu", update_mu, ctx, mu_prior)
  else:
    mu = ctx.rv("mu", mu_prior)
  
  sigma = ctx.rv("sigma", IG(a, b))
  ctx.rv("y", Normal(mu, sigma), obs=y)

If you want to apply an inference algorithm to different subsets of parameters, then a graph and lazy evaluations are needed. The syntax of the PPL would need to be changed.

Something like this:

def model(ctx: Context, y):
  mu = ctx.rv("mu", Normal, 0, 3)
  sigma = ctx.rv("sigma", LogNormal, 0, 1)
  ctx.rv("y", Normal, mu, sigma, obs=y)

rv would have to return a node on a graph.