Allow continuous distributions to be serialized

[mhtess]

For example, is something like this ever conceivable?

Infer({method: "enumerate"}, function(){
    return flip(0.5) ? Beta({a:1,b:1}) : Beta({a:0.01,b:0.01})
})

[stuhlmueller]

Looks fine to me. The code below seems to work:

var hyperDist = Infer({method: "enumerate"}, function(){
  return flip(.5) ? Beta({a:1,b:1}) : Beta({a:0.01,b:0.01})
});

expectation(Infer({method: "rejection", samples: 5000}, function(){
  var dist = sample(hyperDist);
  return sample(dist);
}));

Are you looking for something else?

[mhtess]

Hm. I get not implemented

[stuhlmueller]

Oh, you're right—I wasn't running the test on the public version of webppl. It would definitely be possible to make this work, e.g. by providing a toJSON function for continuous distributions.

Note that this already works:

Infer({method: "enumerate"}, function(){
  return flip(.5) ? Bernoulli({p: .5}) : Bernoulli({p: .7});
});

[longouyang]

Maybe we could serialize continuous distributions (and, say, variational programs; #27) as computation graphs in JSON.