Distributions

Alan wraps all the distributions in torch.distributions, so for all the distributions in torch.distributions, we have a corresponding distribution with exactly the same signature.

The only different thing in Alan is that distributions can take a number of things as arguments:

• A number / tensor (representing a fixed parameter)

• A string (usually representing another random variable)

• A function (usually representing a transformation of another random variable)

• OptParam or QEMParam (representing a learned parameter)

Critically, the string, or __arguments__ to the function must be:

• A previously sampled random variable.

• An input provided to BoundPlate.

• An extra_opt_param provided to BoundPlate.