Finding the Transform Used to Implement Param Constraints

numpyro
1

I’d like to know which exact transformation is used to enforce the constraints for pyro params. For example, below I define the positive parameters of a beta. I’d like to know if, e.g., an exp transform or softplus transform is used.

import pyro
import torch

def guide_beta(x=None):
    # Define the parameters, their initial values and their constraints.
    a0 = pyro.param(
        "a0", 
        torch.tensor(1.0),
        constraint=torch.distributions.constraints.positive
    )

    b0 = pyro.param(
        "b0", 
        torch.tensor(1.0), 
        constraint=torch.distributions.constraints.positive
    )

    # Sample the latent variable using those parameters.
    prob = pyro.sample(
        "prob", 
        pyro.distributions.Beta(a0, b0)
    )
2

look in the relevant pytorch or pyro code for details. e.g. positivity is enforced with an exponential transform by default

3

Oh, excellent, totally missed that file. Thanks!