Probability Distributions

Probability Distributions

Built-In Distributions

Gen.bernoulli — Constant.
bernoulli(prob_true::Real)

Samples a Bool value which is true with given probability

source
Gen.normal — Constant.
normal(mu::Real, std::Real)

Samples a Float64 value from a normal distribution.

source
Gen.mvnormal — Constant.
mvnormal(mu::AbstractVector{T}, cov::AbstractMatrix{U}} where {T<:Real,U<:Real}

Samples a Vector{Float64} value from a multivariate normal distribution.

source
Gen.gamma — Constant.
gamma(shape::Real, scale::Real)

Sample a Float64 from a gamma distribution.

source
Gen.inv_gamma — Constant.
inv_gamma(shape::Real, scale::Real)

Sample a Float64 from a inverse gamma distribution.

source
Gen.beta — Constant.
beta(alpha::Real, beta::Real)

Sample a Float64 from a beta distribution.

source
Gen.categorical — Constant.
categorical(probs::AbstractArray{U, 1}) where {U <: Real}

Given a vector of probabilities probs where sum(probs) = 1, sample an Int i from the set {1, 2, .., length(probs)} with probability probs[i].

source
Gen.uniform — Constant.
uniform(low::Real, high::Real)

Sample a Float64 from the uniform distribution on the interval [low, high].

source
Gen.uniform_discrete — Constant.
uniform_discrete(low::Integer, high::Integer)

Sample an Int from the uniform distribution on the set {low, low + 1, ..., high-1, high}.

source
Gen.poisson — Constant.
poisson(lambda::Real)

Sample an Int from the Poisson distribution with rate lambda.

source
Gen.piecewise_uniform — Constant.
piecewise_uniform(bounds, probs)

Samples a Float64 value from a piecewise uniform continuous distribution.

There are n bins where n = length(probs) and n + 1 = length(bounds). Bounds must satisfy bounds[i] < bounds[i+1] for all i. The probability density at x is zero if x <= bounds[1] or x >= bounds[end] and is otherwise probs[bin] / (bounds[bin] - bounds[bin+1]) where bounds[bin] < x <= bounds[bin+1].

source
Gen.beta_uniform — Constant.
beta_uniform(theta::Real, alpha::Real, beta::Real)

Samples a Float64 value from a mixture of a uniform distribution on [0, 1] with probability 1-theta and a beta distribution with parameters alpha and beta with probability theta.

source

Defining New Distributions

Probability distributions are singleton types whose supertype is Distribution{T}, where T indicates the data type of the random sample.

abstract type Distribution{T} end

By convention, distributions have a global constant lower-case name for the singleton value. For example:

struct Bernoulli <: Distribution{Bool} end
const bernoulli = Bernoulli()

Distributions must implement two methods, random and logpdf.

random returns a random sample from the distribution:

x::Bool = random(bernoulli, 0.5)
x::Bool = random(Bernoulli(), 0.5)

logpdf returns the log probability (density) of the distribution at a given value:

logpdf(bernoulli, false, 0.5)
logpdf(Bernoulli(), false, 0.5)

Distribution values should also be callable, which is a syntactic sugar with the same behavior of calling random:

bernoulli(0.5) # identical to random(bernoulli, 0.5) and random(Bernoulli(), 0.5)

A new Distribution type must implement the following methods:

Gen.random — Function.
val::T = random(dist::Distribution{T}, args...)

Sample a random choice from the given distribution with the given arguments.

source
Gen.logpdf — Function.
lpdf = logpdf(dist::Distribution{T}, value::T, args...)

Evaluate the log probability (density) of the value.

source
Gen.has_output_grad — Function.
has::Bool = has_output_grad(dist::Distribution)

Return true of the gradient if the distribution computes the gradient of the logpdf with respect to the value of the random choice.

source
Gen.logpdf_grad — Function.
grads::Tuple = logpdf_grad(dist::Distribution{T}, value::T, args...)

Compute the gradient of the logpdf with respect to the value, and each of the arguments.

If has_output_grad returns false, then the first element of the returned tuple is nothing. Otherwise, the first element of the tuple is the gradient with respect to the value. If the return value of has_argument_grads has a false value for at position i, then the i+1th element of the returned tuple has value nothing. Otherwise, this element contains the gradient with respect to the ith argument.

source

A new Distribution type must also implement has_argument_grads.