Skip to content

JuliaGaussianProcesses/BayesianLinearRegressors.jl

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

70 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Bayesian Linear Regression in Julia

Build Status Codecov

This is a simple package that does one thing, Bayesian Linear Regression, in around 100 lines of code.

It is actively maintained, but it might appear inactive as it's one of those packages which requires very little maintenance because it's very simple.

Intended Use and Functionality

The interface sits at roughly the same level as that of Distributions.jl. This means that while you won't find a scikit-learn-style fit function, you will find all of the primitives that you need to construct such a function to suit your particular problem. In particular, one can:

  • Construct a BayesianLinearRegressor (BLR) object by providing a mean-vector and precision matrix for the weights of said regressor. This object represents a distribution over (linear) functions.
  • Use a BayesianLinearRegressor as an AbstractGP, as it implements the primary AbstractGP API.
  • Think of an instance of BayesianLinearRegressor as a very restricted GP, where the time complexity of inference scales linearly in the number of observations N.
  • Draw function samples from a BayesianLinearRegressor using rand.
  • Construct a BasisFunctionRegressor object which is a thin wrapper around a BayesianLinearRegressor to allow a non-linear feature mapping ϕ to act on the input.

Conventions

BayesianLinearRegressors is consistent with AbstractGPs. Consequently, a BayesianLinearRegressor in D dimensions can work with the following input types:

  1. ColVecs -- a wrapper around an D x N matrix of Reals saying that each column should be interpreted as an input.
  2. RowVecss -- a wrapper around an N x D matrix of Reals, saying that each row should be interpreted as an input.
  3. Matrix{<:Real} -- must be D x N. Prefer using ColVecs or RowVecs for the sake of being explicit.

Consult the Design section of the KernelFunctions.jl docs for more info on these conventions.

Outputs for a BayesianLinearRegressor should be an AbstractVector{<:Real} of length N.

Example Usage

# Install the packages if you don't already have them installed
] add AbstractGPs BayesianLinearRegressors LinearAlgebra Random Plots Zygote
using AbstractGPs, BayesianLinearRegressors, LinearAlgebra, Random, Plots, Zygote

# Fix seed for re-producibility.
rng = MersenneTwister(123456)

# Construct a BayesianLinearRegressor prior over linear functions of `X`.
mw, Λw = zeros(2), Diagonal(ones(2))
f = BayesianLinearRegressor(mw, Λw)

# Index into the regressor and assume heteroscedastic observation noise `Σ_noise`.
N = 10
X = ColVecs(hcat(range(-5.0, 5.0, length=N), ones(N))')
Σ_noise = Diagonal(exp.(randn(N)))
fX = f(X, Σ_noise)

# Generate some toy data by sampling from the prior.
y = rand(rng, fX)

# Compute the adjoint of `rand` w.r.t. everything given random sensitivities of y′.
_, back_rand = Zygote.pullback(
    (X, Σ_noise, mw, Λw)->rand(rng, BayesianLinearRegressor(mw, Λw)(X, Σ_noise), 5),
    X, Σ_noise, mw, Λw,
)
back_rand(randn(N, 5))

# Compute the `logpdf`. Read as `the log probability of observing `y` at `X` under `f`, and
# Gaussian observation noise with zero-mean and covariance `Σ_noise`.
logpdf(fX, y)

# Compute the gradient of the `logpdf` w.r.t. everything.
Zygote.gradient(
    (X, Σ_noise, y, mw, Λw)->logpdf(BayesianLinearRegressor(mw, Λw)(X, Σ_noise), y),
    X, Σ_noise, y, mw, Λw,
)

# Perform posterior inference. Note that `f′` has the same type as `f`.
f′ = posterior(fX, y)

# Compute `logpdf` of the observations under the posterior predictive.
logpdf(f′(X, Σ_noise), y)

# Sample from the posterior predictive distribution.
N_plt = 1000
X_plt = ColVecs(hcat(range(-6.0, 6.0, length=N_plt), ones(N_plt))')

# Compute some posterior marginal statisics.
normals = marginals(f′(X_plt, eps()))
m′X_plt = mean.(normals)
σ′X_plt = std.(normals)

# Plot the posterior. This uses the default AbstractGPs plotting recipes.
posterior_plot = plot();
plot!(posterior_plot, X_plt.X[1, :], f′(X_plt, eps()); color=:blue, ribbon_scale=3);
sampleplot!(posterior_plot, X_plt.X[1, :], f′(X_plt, eps()); color=:blue, samples=10);
scatter!(posterior_plot, X.X[1, :], y; # Observations.
    markercolor=:red,
    markershape=:circle,
    markerstrokewidth=0.0,
    markersize=4,
    markeralpha=0.7,
    label="",
);
display(posterior_plot);

Basis Function Regression

Any instance of a BayesianLinearRegressor can be replaced by a BasisFunctionRegressor (BFR). A BasisFunctionRegressor is a thin wrapper around a BayesianLinearRegressor, but includes a potentially non-linear feature mapping ϕ which is applied to the input before it is passed to the underlying BLR. It is essentially defined as bfr(X) = blr(ϕ(X)).

using AbstractGPs, BayesianLinearRegressors, LinearAlgebra

X = RowVecs(hcat(range(-1.0, 1.0, length=5)))
blr = BayesianLinearRegressor(zeros(2), Diagonal(ones(2)))

# N.B. ϕ must accept one of the allowed input types and
# must return the same type (in this case RowVecs)
ϕ(x::RowVecs) = RowVecs(hcat(ones(length(x)), prod.(x)))

bfr = BasisFunctionRegressor(blr, ϕ)

# These are equivalent
var(bfr(X)) == var(blr(ϕ(X)))

Drawing Function Samples

There are two ways of drawing samples from f::Union{BayesianLinearRegressor,BasisFunctionRegressor}. The first is by using the AbstractGPs API (as in the above examples) where a FiniteBLR projection is first produced at a fixed set of input locations X with some specified observation noise Σy: fx = f(X, Σy)::FiniteBLR. Then, observations (including noise) at X can be sampled directly with rand(rng, fx). The other way is to draw a sample of an entire function from f using g = rand(rng, f). This samples a value for the weights w ~ N(mw, Λw) and produces a function g(X) = w'X which can be evaluated at any input locations. Note that this method corresponds to drawing samples of noiseless observations.

using AbstractGPs, BayesianLinearRegressors, LinearAlgebra, Random, Plots

# Fix seed for re-producibility.
rng = MersenneTwister(123456)

X = RowVecs(hcat(range(-1.0, 1.0, length=5)))
f = BayesianLinearRegressor(zeros(2), Diagonal(ones(2)))

## The first method of drawing samples - using the AbstractGPs API:
# Index into the regressor at fixed inputs X and assume homoscedastic observation noise `Σ_noise`.
N = 10
X = ColVecs(hcat(range(-5.0, 5.0, length=N), ones(N))')
Σ_noise = 0.1
fX = f(X, Σ_noise)

rand(rng, fX)

## The second method - sampling an entire function:
g = rand(rng, f)

# This sample can now be evaulated at any input locations
g(X)

X′ = ColVecs(hcat(range(10.0, 15.0, length=N), ones(N))')
g(X′)

# Sample multiple functions for plotting
gs = rand(rng, f, 100)

N_plt = 1000
X_plt = ColVecs(hcat(range(-6.0, 6.0, length=N_plt), ones(N_plt))')
y_plt = reduce(hcat, [g(X_plt) for g in gs])

sample_plot = plot();
plot!(sample_plot, X_plt.X[1,:], y_plt;
    label="",
    color=:black,
    linealpha=0.4,
);
display(sample_plot)

Up For Grabs

  • Scikit-learn style interface: it wouldn't be too hard to implement a scikit-learn - style interface to handle basic regression tasks, so please feel free to make a PR that implements this.
  • Monte Carlo VI (MCVI): i.e. variational inference using the reparametrisation trick. This could be very useful when working with large data sets and applying big non-linear transformations, such as neural networks, to the inputs as it would enable mini-batching. I would envisage at least supporting both a dense approximate posterior covariance and diagonal (i.e. mean-field), where the former is for small-moderate dimensionalities and the latter for very high-dimensional problems.

Bugs, Issues, and PRs

Please do report any bugs you find by raising an issue. Please also feel free to raise PRs, especially if for one of the above Up For Grabs items. Raise an issue to discuss the extension in detail before opening a PR if you prefer, though.

Related Work

BayesianLinearRegression.jl is closely related, but appears to be a WIP and hasn't been touched in around a year or so (as of 27-03-2019).