Using transforms to enforce positivity of sum of RVs

Misc.
1

Hi all,

pyro newbie here. I’m trying to figure out how to use transformations (in particular, softplus) to enforce the positivity of a sum of variables. My current attempt below returns an “AttributeError: ‘Tensor’ object has no attribute ‘batch_shape’”, because a+b is apparently a Tensor and not a distribution.

Any suggestions on how to fix this? Alternative approaches to achieve the same result (constraining the sum of a and b to be positive) are also very welcome. :slight_smile:

(What I actually want to do is sample from a Bernoulli distribution, with the probability given by the sum of various other RVs. Thus I need to somehow ensure that that sum is between 0 and 1. Choosing constraints on all summands such that the sum is always within [0,1] is extremely restrictive.)

import pyro
import pyro.distributions as dist
from pyro.infer import Predictive
from pyro.distributions.transforms import SoftplusTransform

def model():
    a = pyro.sample("a", dist.Uniform(0.0, 1.0))
    b = pyro.sample("b", dist.Uniform(-0.1, 0.1))
    var = pyro.sample("var",  dist.TransformedDistribution(a+b, transforms=[SoftplusTransform()]))

# draw samples from the prior
data = Predictive(model, {}, num_samples=1000)()
print(data)
2

I figured out how to do what I want. I’ll leave this here for future reference.

import pyro
import pyro.distributions as dist
from pyro.infer import Predictive
import torch

def model():
    a = pyro.sample("a", dist.Uniform(0.0, 1.0))
    b = pyro.sample("b", dist.Uniform(-0.1, 0.1))
    var = pyro.deterministic("var", torch.nn.functional.softplus(a+b))

# draw samples from the prior
data = Predictive(model, {}, num_samples=1000)()
print(data)
3 Likes