Specifying model by (unnormalized) joint density

Misc.
1

Hi everyone,

is it possible in pyro to define the model not by pyro.sample statements but instead directly define it by an (unnormalized) joint density.

Thanks,
Florian

1 Like
2

Usually I am able to specify a joint density via pyro.sample(..., obs=...) statements, i.e. via a factor graph representation. If the standard distributions are insufficient, you can wrap an arbitrary density in a TorchDistribution object:

class MyCustomDistribution(pyro.distributions.TorchDistribution):
    def __init__(self, ...):
        super(MyCustomDistribution, self)(...)
    def log_prob(self, value):
        return my_custom_log_density(value)

then in a Pyro model you can call

def model(data):
    pyro.sample("density", MyCustomDistribution(...), obs=data)

Note that since you haven’t defined a .sample() method, this custom density can only be used in the model, not in the guide.

3 Likes
3

Thanks, that’s exactly what I needed. :slight_smile: