transform(b, x)

Transform x using b, treating x as a single input.

logabsdetjac(b, x)

Return log(abs(det(J(b, x)))), where J(b, x) is the jacobian of b at x.

logabsdetjacinv(b, y)

Just an alias for logabsdetjac(inverse(b), y).

inverse(t::Transform)

Returns the inverse of transform t.

transform!(b, x[, y])

Transform x using b, storing the result in y.

If y is not provided, x is used as the output.

logabsdetjac!(b, x[, logjac])

Compute log(abs(det(J(b, x)))) and store the result in logjac, where J(b, x) is the jacobian of b at x.

with_logabsdet_jacobian!(b, x[, y, logjac])

Compute transform(b, x) and logabsdetjac(b, x), storing the result in y and logjac, respetively.

If y is not provided, then x will be used in its place.

Defaults to calling with_logabsdet_jacobian(b, x) and updating y and logjac with the result.

bijector(d::Distribution)

Returns the constrained-to-unconstrained bijector for distribution d.

transformed(d::Distribution)
transformed(d::Distribution, b::Bijector)

Couples distribution d with the bijector b by returning a TransformedDistribution.

If no bijector is provided, i.e. transformed(d) is called, then transformed(d, bijector(d)) is returned.

output_size(f, sz)

Returns the output size of f given the input size sz.

output_size(f, dist::Distribution)

Returns the output size of f given the input distribution dist. This is useful when Base.size(dist) is not defined, e.g. for ProductNamedTupleDistribution and in particular is used by DynamicPPL when generating new random values for transformed distributions.

By default this just calls output_size(f, size(dist)), but this can be overloaded for specific distributions.

elementwise(f)

Alias for Base.Fix1(broadcast, f).

In the case where f::ComposedFunction, the result is Base.Fix1(broadcast, f.outer) ∘ Base.Fix1(broadcast, f.inner) rather than Base.Fix1(broadcast, f).

isinvertible(t)

Return true if t is invertible, and false otherwise.

isclosedform(b::Transform)::bool
isclosedform(b⁻¹::Inverse{<:Transform})::bool

Returns true or false depending on whether or not evaluation of b has a closed-form implementation.

Most transformations have closed-form evaluations, but there are cases where this is not the case. For example the inverse evaluation of PlanarLayer requires an iterative procedure to evaluate.

Abstract type for a transformation.

Implementing

A subtype of Transform of should at least implement transform(b, x).

If the Transform is also invertible:

• Required:
  • Either of the following:
    • transform(::Inverse{<:MyTransform}, x): the transform for its inverse.
    • InverseFunctions.inverse(b::MyTransform): returns an existing Transform.
  • logabsdetjac: computes the log-abs-det jacobian factor.
• Optional:
  • with_logabsdet_jacobian: transform and logabsdetjac combined. Useful in cases where we can exploit shared computation in the two.

For the above methods, there are mutating versions which can optionally be implemented:

• with_logabsdet_jacobian!
• logabsdetjac!
• with_logabsdet_jacobian!

Abstract type of a bijector, i.e. differentiable bijection with differentiable inverse.

inverse(b::Transform)
Inverse(b::Transform)

A Transform representing the inverse transform of b.

CorrBijector <: Bijector

A bijector implementation of Stan's parametrization method for Correlation matrix: https://mc-stan.org/docs/reference-manual/transforms.html#correlation-matrix-transform.section

Basically, a unconstrained strictly upper triangular matrix y is transformed to a correlation matrix by following readable but not that efficient form:

K = size(y, 1)
z = tanh.(y)

for j=1:K, i=1:K
    if i>j
        w[i,j] = 0
    elseif 1==i==j
        w[i,j] = 1
    elseif 1<i==j
        w[i,j] = prod(sqrt(1 .- z[1:i-1, j].^2))
    elseif 1==i<j
        w[i,j] = z[i,j]
    elseif 1<i<j
        w[i,j] = z[i,j] * prod(sqrt(1 .- z[1:i-1, j].^2))
    end
end

It is easy to see that every column is a unit vector, for example:

w3' w3 ==
w[1,3]^2 + w[2,3]^2 + w[3,3]^2 ==
z[1,3]^2 + (z[2,3] * sqrt(1 - z[1,3]^2))^2 + (sqrt(1-z[1,3]^2) * sqrt(1-z[2,3]^2))^2 ==
z[1,3]^2 + z[2,3]^2 * (1-z[1,3]^2) + (1-z[1,3]^2) * (1-z[2,3]^2) ==
z[1,3]^2 + z[2,3]^2 - z[2,3]^2 * z[1,3]^2 + 1 -z[1,3]^2 - z[2,3]^2 + z[1,3]^2 * z[2,3]^2 ==
1

And diagonal elements are positive, so w is a cholesky factor for a positive matrix.

x = w' * w

Consider block matrix representation for x

x = [w1'; w2'; ... wn'] * [w1 w2 ... wn] == 
[w1'w1 w1'w2 ... w1'wn;
 w2'w1 w2'w2 ... w2'wn;
 ...
]

The diagonal elements are given by wk'wk = 1, thus x is a correlation matrix.

Every step is invertible, so this is a bijection(bijector).

Note: The implementation doesn't follow their "manageable expression" directly, because their equation seems wrong (7/30/2020). Insteadly it follows definition above the "manageable expression" directly, which is also described in above doc.

LeakyReLU{T}(α::T) <: Bijector

Defines the invertible mapping

x ↦ x if x ≥ 0 else αx

where α > 0.

Stacked(bs)
Stacked(bs, ranges)
Stacked(bs::Bijector...)

A Bijector which stacks bijectors together which can then be applied to a vector where bs[i]::Bijector is applied to x[ranges[i]]::UnitRange{Int}.

Arguments

• bs can be either a Tuple or an AbstractArray of 0- and/or 1-dimensional bijectors
  • If bs is a Tuple, implementations are type-stable using generated functions
  • If bs is an AbstractArray, implementations are not type-stable and use iterative methods
• ranges needs to be an iterable consisting of UnitRange{Int}
  • length(bs) == length(ranges) needs to be true.

Examples

using Bijectors: Logit, Stacked
b1 = Logit(0.0, 1.0)
b2 = identity
b = Stacked(b1, b2)
b([0.0, 1.0]) == [b1(0.0), 1.0]  # => true

RationalQuadraticSpline{T} <: Bijector

Implementation of the Rational Quadratic Spline flow [1].

• Outside of the interval [minimum(widths), maximum(widths)], this mapping is given by the identity map.
• Inside the interval it's given by a monotonic spline (i.e. monotonic polynomials connected at intermediate points) with endpoints fixed so as to continuously transform into the identity map.

For the sake of efficiency, there are separate implementations for 0-dimensional and 1-dimensional inputs.

Notes

There are two constructors for RationalQuadraticSpline:

• RationalQuadraticSpline(widths, heights, derivatives): it is assumed that widths,

heights, and derivatives satisfy the constraints that makes this a valid bijector, i.e.

• widths: monotonically increasing and length(widths) == K,
• heights: monotonically increasing and length(heights) == K,
• derivatives: non-negative and derivatives[1] == derivatives[end] == 1.
• RationalQuadraticSpline(widths, heights, derivatives, B): other than than the lengths, no assumptions are made on parameters. Therefore we will transform the parameters s.t.:
• widths_new ∈ [-B, B]ᴷ⁺¹, where K == length(widths),
• heights_new ∈ [-B, B]ᴷ⁺¹, where K == length(heights),
• derivatives_new ∈ (0, ∞)ᴷ⁺¹ with derivatives_new[1] == derivates_new[end] == 1, where (K - 1) == length(derivatives).

Examples

Univariate

julia> using StableRNGs: StableRNG; rng = StableRNG(42);  # For reproducibility.

julia> using Bijectors: RationalQuadraticSpline

julia> K = 3; B = 2;

julia> # Monotonic spline on '[-B, B]' with `K` intermediate knots/"connection points".
       b = RationalQuadraticSpline(randn(rng, K), randn(rng, K), randn(rng, K - 1), B);

julia> b(0.5) # inside of `[-B, B]` → transformed
1.1943325397834206

julia> b(5.) # outside of `[-B, B]` → not transformed
5.0

julia> b = RationalQuadraticSpline(b.widths, b.heights, b.derivatives);

julia> b(0.5) # inside of `[-B, B]` → transformed
1.1943325397834206

julia> d = 2; K = 3; B = 2;

julia> b = RationalQuadraticSpline(randn(rng, d, K), randn(rng, d, K), randn(rng, d, K - 1), B);

julia> b([-1., 1.])
2-element Vector{Float64}:
 -1.5660106244288925
  0.5384702734738573

julia> b([-5., 5.])
2-element Vector{Float64}:
 -5.0
  5.0

julia> b([-1., 5.])
2-element Vector{Float64}:
 -1.5660106244288925
  5.0

References

[1] Durkan, C., Bekasov, A., Murray, I., & Papamakarios, G., Neural Spline Flows, CoRR, arXiv:1906.04032 [stat.ML], (2019).

Coupling{F, M}(θ::F, mask::M)

Implements a coupling-layer as defined in [1].

Examples

julia> using Bijectors: Shift, Coupling, PartitionMask, coupling, couple

julia> m = PartitionMask(3, [1], [2]); # <= going to use x[2] to parameterize transform of x[1]

julia> cl = Coupling(Shift, m); # <= will do `y[1:1] = x[1:1] + x[2:2]`;

julia> x = [1., 2., 3.];

julia> cl(x)
3-element Vector{Float64}:
 3.0
 2.0
 3.0

julia> inverse(cl)(cl(x))
3-element Vector{Float64}:
 1.0
 2.0
 3.0

julia> coupling(cl) # get the `Bijector` map `θ -> b(⋅, θ)`
Shift

julia> couple(cl, x) # get the `Bijector` resulting from `x`
Shift{Vector{Float64}}([2.0])

julia> with_logabsdet_jacobian(cl, x)
([3.0, 2.0, 3.0], 0.0)

References

[1] Kobyzev, I., Prince, S., & Brubaker, M. A., Normalizing flows: introduction and ideas, CoRR, (), (2019).

OrderedBijector()

A bijector mapping unordered vectors in ℝᵈ to ordered vectors in ℝᵈ.

See also

• Stan's documentation
  • Note that this transformation and its inverse are the opposite of in this reference.

NamedTransform <: AbstractNamedTransform

Wraps a NamedTuple of key -> Bijector pairs, implementing evaluation, inversion, etc.

Examples

julia> using Bijectors: NamedTransform, Scale

julia> b = NamedTransform((a = Scale(2.0), b = exp));

julia> x = (a = 1., b = 0., c = 42.);

julia> b(x)
(a = 2.0, b = 1.0, c = 42.0)

julia> (a = 2 * x.a, b = exp(x.b), c = x.c)
(a = 2.0, b = 1.0, c = 42.0)

NamedCoupling{target, deps, F} <: AbstractNamedTransform

Implements a coupling layer for named bijectors.

See also: Coupling

Examples

julia> using Bijectors: NamedCoupling, Scale

julia> b = NamedCoupling(:b, (:a, :c), (a, c) -> Scale(a + c));

julia> x = (a = 1., b = 2., c = 3.);

julia> b(x)
(a = 1.0, b = 8.0, c = 3.0)

julia> (a = x.a, b = (x.a + x.c) * x.b, c = x.c)
(a = 1.0, b = 8.0, c = 3.0)