Getting started

RxInfer.jl is a Julia package for Bayesian Inference on Factor Graphs by Message Passing. It supports both exact and variational inference algorithms and forms an ecosystem around three main packages:

• ReactiveMP.jl - the underlying message passing-based inference engine
• GraphPPL.jl - model and constraints specification package
• Rocket.jl - reactive extensions package for Julia

This page provides the necessary information you need to get started with Rxinfer. We will show the general approach to solving inference problems with RxInfer by means of a running example: inferring the bias of a coin using a simple Beta-Bernoulli model.

Installation

RxInfer is an officially registered Julia package. Install RxInfer through the Julia package manager by using the following command from the package manager mode:

] add RxInfer

Alternatively:

julia> import Pkg

julia> Pkg.add("RxInfer")

Importing RxInfer

To add RxInfer package (and all associated packages) into a running Julia session simply run:

using RxInfer

Read more about about using in the Using methods from RxInfer section of the documentation.

Example: Inferring the bias of a coin

The RxInfer approach to solving inference problems consists of three phases:

1. Model specification: RxInfer uses GraphPPL package for model specification part. It offers a domain-specific language to specify your probabilistic model.
2. Inference specification: RxInfer inference API uses ReactiveMP inference engine under the hood and has been designed to be as flexible as possible. It is compatible both with asynchronous infinite data streams and with static datasets. For most of the use cases it consists of the same simple building blocks. In this example we will show one of the many possible ways to infer your quantities of interest.
3. Inference execution: Given model specification and inference procedure it is pretty straightforward to use package's API to pass data to the inference backend and to run actual inference.

Coin flip simulation

Let's start by creating some dataset. One approach could be flipping a coin N times and recording each outcome. For simplicity in this example we will use static pre-generated dataset. Each sample can be thought of as the outcome of single flip which is either heads or tails (1 or 0). We will assume that our virtual coin is biased, and lands heads up on 75% of the trials (on average).

First let's setup our environment by importing all needed packages:

using RxInfer, Distributions, Random

Next, let's define our dataset:

# Random number generator for reproducibility
rng            = MersenneTwister(42)
# Number of coin flips (observations)
n_observations = 10
# The bias of a coin used in the demonstration
coin_bias      = 0.75
# We assume that the outcome of each coin flip is
# distributed as the `Bernoulli` distrinution
distribution   = Bernoulli(coin_bias)
# Simulated coin flips
dataset        = rand(rng, distribution, n_observations)

10-element Vector{Bool}:
 1
 1
 1
 1
 0
 0
 1
 1
 0
 1

Model specification

In a Bayesian setting, the next step is to specify our probabilistic model. This amounts to specifying the joint probability of the random variables of the system.

Likelihood

We've made an assumption that the outcome of each coin flip is governed by the Bernoulli distribution, i.e.

\[y_i \sim \mathrm{Bernoulli}(\theta),\]

where $y_i = 1$ represents "heads", $y_i = 0$ represents "tails". The underlying probability of the coin landing heads up for a single coin flip is $\theta \in [0,1]$.

Prior

We will choose the conjugate prior of the Bernoulli likelihood function defined above, namely the Beta distribution, i.e.

\[\theta \sim Beta(a, b),\]

where $a$ and $b$ are the hyperparameters that encode our prior beliefs about the possible values of $\theta$. We will assign values to the hyperparameters in a later step.

Joint probability

The joint probability is given by the multiplication of the likelihood and the prior, i.e.

\[P(y_{1:N}, θ) = P(θ) \prod_{i=1}^N P(y_i | θ).\]

Now let's see how to specify this model using GraphPPL's package syntax.

# GraphPPL.jl export `@model` macro for model specification
# It accepts a regular Julia function and builds an FFG under the hood
@model function coin_model(y, a, b)
    # We endow θ parameter of our model with some prior
    θ ~ Beta(a, b)
    # or, in this particular case, the `Uniform(0.0, 1.0)` prior also works:
    # θ ~ Uniform(0.0, 1.0)

    # We assume that outcome of each coin flip is governed by the Bernoulli distribution
    for i in eachindex(y)
        y[i] ~ Bernoulli(θ)
    end
end

As you can see, RxInfer offers a model specification syntax that resembles closely to the mathematical equations defined above. Alternatively, we could use a broadcasting syntax:

@model function coin_model(y, a, b)
    θ  ~ Beta(a, b)
    y .~ Bernoulli(θ)
end

Conditioning on data and inspecting the model structure

Given the model specification we can construct an actual model graph and visualize it. In order to do that we can use the | operator to condition on data and the RxInfer.create_model function to create the graph. Read more about condition in the corresponding section of the documentation.

conditioned = coin_model(a = 2.0, b = 7.0) | (y = [ true, false, true ], )

coin_model(a = 2.0, b = 7.0) conditioned on: 
  y = Bool[1, 0, 1]

We can use GraphPPL.jl visualisation capabilities to show the structure of the resulting graph. For that we need two extra packages installed: Cairo and GraphPlot. Note, that those packages are not included in the RxInfer package and must be installed separately.

using Cairo, GraphPlot

# `Create` the actual graph of the model conditioned on the data
model = RxInfer.create_model(conditioned)

# Call `gplot` function from `GraphPlot` to visualise the structure of the graph
GraphPlot.gplot(RxInfer.getmodel(model))

In addition, we can also programatically query the structure of the graph:

RxInfer.getrandomvars(model)

1-element Vector{GraphPPL.NodeData}:
 NodeData in context  with properties name = θ, index = nothing

RxInfer.getdatavars(model)

3-element Vector{GraphPPL.NodeData}:
 NodeData in context  with properties name = y, index = 1
 NodeData in context  with properties name = y, index = 2
 NodeData in context  with properties name = y, index = 3

RxInfer.getconstantvars(model)

2-element Vector{GraphPPL.NodeData}:
 NodeData in context  with properties name = constvar_2, index = nothing
 NodeData in context  with properties name = constvar_4, index = nothing

RxInfer.getfactornodes(model)

4-element Vector{GraphPPL.NodeData}:
 NodeData in context  with properties fform = Beta, neighbors = Tuple{GraphPPL.NodeLabel, GraphPPL.EdgeLabel, GraphPPL.NodeData}[(θ_1, out, NodeData in context  with properties name = θ, index = nothing), (constvar_2_3, a, NodeData in context  with properties name = constvar_2, index = nothing), (constvar_4_5, b, NodeData in context  with properties name = constvar_4, index = nothing)]
 NodeData in context  with properties fform = Bernoulli, neighbors = Tuple{GraphPPL.NodeLabel, GraphPPL.EdgeLabel, GraphPPL.NodeData}[(y_7, out, NodeData in context  with properties name = y, index = 1), (θ_1, p, NodeData in context  with properties name = θ, index = nothing)]
 NodeData in context  with properties fform = Bernoulli, neighbors = Tuple{GraphPPL.NodeLabel, GraphPPL.EdgeLabel, GraphPPL.NodeData}[(y_8, out, NodeData in context  with properties name = y, index = 2), (θ_1, p, NodeData in context  with properties name = θ, index = nothing)]
 NodeData in context  with properties fform = Bernoulli, neighbors = Tuple{GraphPPL.NodeLabel, GraphPPL.EdgeLabel, GraphPPL.NodeData}[(y_9, out, NodeData in context  with properties name = y, index = 3), (θ_1, p, NodeData in context  with properties name = θ, index = nothing)]

Conditioning on data that is not available at model creation time

Sometimes the data is not known at model creation time, for example, during reactive inference. For that purpose RxInfer uses RxInfer.DeferredDataHandler structure.

# The only difference here is that we do not specify `a` and `b` as hyperparameters
# But rather indicate that the data for them will be available later during the inference
conditioned_with_deffered_data = coin_model() | (
    y = [ true, false, true ],
    a = RxInfer.DeferredDataHandler(),
    b = RxInfer.DeferredDataHandler()
)

# The graph creation API does not change
model_with_deffered_data = RxInfer.create_model(conditioned_with_deffered_data)

# We can visualise the graph with missing data handles as well
GraphPlot.gplot(RxInfer.getmodel(model_with_deffered_data))

From the model structure visualisation we can see now that both a and b are no longer indicated as constants. Read more about the structure of the graph in GraphPPL documentation.

Inference specification

Automatic inference specification

Once we have defined our model, the next step is to use RxInfer API to infer quantities of interests. To do this we can use a generic infer function that supports static datasets. Read more information about the infer function in the Inference Execution documentation section.

result = infer(
    model = coin_model(a = 2.0, b = 7.0),
    data  = (y = dataset, )
)

Inference results:
  Posteriors       | available for (θ)

As you can see we don't need to condition on the data manually, the infer function will do it automatically. After the inference is complete we can fetch the results from the .posterior field with the name of the latent state:

θestimated = result.posteriors[:θ]

Beta{Float64}(α=9.0, β=10.0)

We can also compute some statistical properties of the result:

println("Real bias is ", coin_bias)
println("Estimated bias is ", mean(θestimated))
println("Standard deviation ", std(θestimated))

Real bias is 0.75
Estimated bias is 0.47368421052631576
Standard deviation 0.11164843913471803

Let's also visualize the resulting posteriors:

using Plots

rθ = range(0, 1, length = 1000)

p1 = plot(rθ, (x) -> pdf(Beta(2.0, 7.0), x), title="Prior", fillalpha=0.3, fillrange = 0, label="P(θ)", c=1,)
p2 = plot(rθ, (x) -> pdf(θestimated, x), title="Posterior", fillalpha=0.3, fillrange = 0, label="P(θ|y)", c=3)

plot(p1, p2, layout = @layout([ a; b ]))

In our dataset we used 10 coin flips and skewed prior to estimate the bias of a coin. It resulted in a vague posterior distribution, however RxInfer scales very well for large models and factor graphs. We may use more coin flips in our dataset for better posterior distribution estimates:

dataset_100   = rand(rng, Bernoulli(coin_bias), 100)
dataset_1000  = rand(rng, Bernoulli(coin_bias), 1000)
dataset_10000 = rand(rng, Bernoulli(coin_bias), 10000)

θestimated_100   = infer(model = coin_model(a = 2.0, b = 7.0), data  = (y = dataset_100, ))
θestimated_1000  = infer(model = coin_model(a = 2.0, b = 7.0), data  = (y = dataset_1000, ))
θestimated_10000 = infer(model = coin_model(a = 2.0, b = 7.0), data  = (y = dataset_10000, ))

Let's investigate how the number of observation affects the estimated posterior:

p3 = plot(title = "Posterior", legend = :topleft)
p3 = plot!(p3, rθ, (x) -> pdf(θestimated_100.posteriors[:θ], x), fillalpha = 0.3, fillrange = 0, label = "P(θ|y_100)", c = 4)
p3 = plot!(p3, rθ, (x) -> pdf(θestimated_1000.posteriors[:θ], x), fillalpha = 0.3, fillrange = 0, label = "P(θ|y_1000)", c = 5)
p3 = plot!(p3, rθ, (x) -> pdf(θestimated_10000.posteriors[:θ], x), fillalpha = 0.3, fillrange = 0, label = "P(θ|y_10000)", c = 6)

plot(p1, p3, layout = @layout([ a; b ]))

We can see that with larger dataset our posterior marginal estimate becomes more and more accurate and represents real value of the bias of a coin.

println("Real bias is ", coin_bias)
println("Estimated bias is ", mean(θestimated_10000.posteriors[:θ]))
println("Standard deviation ", std(θestimated_10000.posteriors[:θ]))

Real bias is 0.75
Estimated bias is 0.7529223698671196
Standard deviation 0.004310967761308469

Missing data points

RxInfer inference engine understands missing data points in the dataset, for example:

result = infer(
    model = coin_model(a = 2.0, b = 7.0),
    data  = (y = [ true, false, missing, true, false ], )
)
result.posteriors[:θ]

Beta{Float64}(α=4.0, β=9.0)

In principle, the entire dataset may consist of missing entries:

result = infer(
    model = coin_model(a = 2.0, b = 7.0),
    data  = (y = [ missing, missing, missing, missing, missing ], )
)

Inference results:
  Posteriors       | available for (θ)
  Predictions      | available for (y)

In this case, the resulting posterior is simply equal to the prior (as expected, since no extra information can be extracted from the observations):

result.posteriors[:θ]

Beta{Float64}(α=2.0, β=7.0)

Where to go next?

There are a set of examples available in RxInfer repository that demonstrate the more advanced features of the package for various problems. Alternatively, you can head to the Model specification which provides more detailed information of how to use RxInfer to specify probabilistic models. Inference execution section provides a documentation about RxInfer API for running reactive Bayesian inference. Also read the Comparison to compare RxInfer with other probabilistic programming libraries.