Querying Model Probabilities

The easiest way to manipulate and query Turing models is via the DynamicPPL probability interface.

Let’s use a simple model of normally-distributed data as an example.

using Turing
using DynamicPPL
using Random

@model function gdemo(n)
    μ ~ Normal(0, 1)
    x ~ MvNormal(fill(μ, n), I)
end
gdemo (generic function with 2 methods)

We generate some data using μ = 0:

Random.seed!(1776)
dataset = randn(100)
dataset[1:5]
5-element Vector{Float64}:
  0.8488780584442736
 -0.31936138249336765
 -1.3982098801744465
 -0.05198933163879332
 -1.1465116601038348

Conditioning and Deconditioning

Bayesian models can be transformed with two main operations, conditioning and deconditioning (also known as marginalization). Conditioning takes a variable and fixes its value as known. We do this by passing a model and a collection of conditioned variables to |, or its alias, condition:

# (equivalently)
# conditioned_model = condition(gdemo(length(dataset)), (x=dataset, μ=0))
conditioned_model = gdemo(length(dataset)) | (x=dataset, μ=0)
Model{typeof(gdemo), (:n,), (), (), Tuple{Int64}, Tuple{}, ConditionContext{@NamedTuple{x::Vector{Float64}, μ::Int64}, DefaultContext}}(gdemo, (n = 100,), NamedTuple(), ConditionContext((x = [0.8488780584442736, -0.31936138249336765, -1.3982098801744465, -0.05198933163879332, -1.1465116601038348, -0.6306168227545849, 0.6862766694322289, -0.5485073478947856, -0.17212004616875684, 1.2883226251958486, -0.13661316034377538, 2.4316115122026973, 0.2251319215717449, -0.5115708179083417, -0.7810712258995324, -1.0191704692490737, 1.1210038448250719, -1.6944509713762377, -0.27314823183454695, 0.25273963222687423, 1.3914215917992434, 0.7525340831125464, 0.847154387311101, -0.7130402796655171, 0.2983575202861233, -0.1785631526879386, 0.08659477535701691, -0.5167265137098563, 2.111309740316035, 0.3957655443124509, -0.0804390853521051, 1.255042471667049, -0.07882822403959532, 1.2261373761992618, 0.43953618247769816, -0.40640013183427787, -0.6868635949523503, 1.7380713294668497, 0.13685965156352295, 0.1485185624825999, -0.7798816720822024, 2.2595105995080846, -0.13609014938597142, 0.22785777205259913, -2.1005250433485725, 0.44205288222935385, -1.238456637875994, -2.3727125492433427, -0.24406624959402184, -0.04488042525902438, 0.27510026183444175, 0.42472846594528796, 1.0337924022589282, 0.9126364433535069, -0.9006583845907805, 0.8665471057463393, 1.4924737539852484, 1.2886591566091432, 1.037264411147446, 1.4731954133339449, -0.31874662373651885, 1.2255399151799211, -1.6642044048811695, -0.5717328092786154, -1.2700237196779645, 0.5748199649058684, 0.16467729820692942, -1.195290550625328, -0.37133526877621703, -0.3018979982049836, -2.0183406292097397, -0.9588803575112745, 0.7177183994733006, -1.0133440177662316, -1.0881357990941283, 1.0487446580734279, 2.627227367991459, -1.59963908284846, -0.3122512299247273, -1.0265333654194488, 0.5557085182114885, -0.3206725445321106, -1.4314746067673778, 1.5740113510560039, -0.6566477752702335, 0.31342313477927125, 0.33135361418686027, -1.0489180508346863, -0.2670759024309527, 0.4683952221006179, 0.04918061587657951, 1.239814741442417, 2.2239462179369296, 1.8507671783064434, 1.756319462015174, -0.6577450354719728, 2.2795431083561626, -0.492273906928334, 0.7045614632761499, 0.11260553216111485], μ = 0), DynamicPPL.DefaultContext()))

This operation can be reversed by applying decondition:

original_model = decondition(conditioned_model)
Model{typeof(gdemo), (:n,), (), (), Tuple{Int64}, Tuple{}, DefaultContext}(gdemo, (n = 100,), NamedTuple(), DefaultContext())

We can also decondition only some of the variables:

partially_conditioned = decondition(conditioned_model, :μ)
Model{typeof(gdemo), (:n,), (), (), Tuple{Int64}, Tuple{}, ConditionContext{@NamedTuple{x::Vector{Float64}}, DefaultContext}}(gdemo, (n = 100,), NamedTuple(), ConditionContext((x = [0.8488780584442736, -0.31936138249336765, -1.3982098801744465, -0.05198933163879332, -1.1465116601038348, -0.6306168227545849, 0.6862766694322289, -0.5485073478947856, -0.17212004616875684, 1.2883226251958486, -0.13661316034377538, 2.4316115122026973, 0.2251319215717449, -0.5115708179083417, -0.7810712258995324, -1.0191704692490737, 1.1210038448250719, -1.6944509713762377, -0.27314823183454695, 0.25273963222687423, 1.3914215917992434, 0.7525340831125464, 0.847154387311101, -0.7130402796655171, 0.2983575202861233, -0.1785631526879386, 0.08659477535701691, -0.5167265137098563, 2.111309740316035, 0.3957655443124509, -0.0804390853521051, 1.255042471667049, -0.07882822403959532, 1.2261373761992618, 0.43953618247769816, -0.40640013183427787, -0.6868635949523503, 1.7380713294668497, 0.13685965156352295, 0.1485185624825999, -0.7798816720822024, 2.2595105995080846, -0.13609014938597142, 0.22785777205259913, -2.1005250433485725, 0.44205288222935385, -1.238456637875994, -2.3727125492433427, -0.24406624959402184, -0.04488042525902438, 0.27510026183444175, 0.42472846594528796, 1.0337924022589282, 0.9126364433535069, -0.9006583845907805, 0.8665471057463393, 1.4924737539852484, 1.2886591566091432, 1.037264411147446, 1.4731954133339449, -0.31874662373651885, 1.2255399151799211, -1.6642044048811695, -0.5717328092786154, -1.2700237196779645, 0.5748199649058684, 0.16467729820692942, -1.195290550625328, -0.37133526877621703, -0.3018979982049836, -2.0183406292097397, -0.9588803575112745, 0.7177183994733006, -1.0133440177662316, -1.0881357990941283, 1.0487446580734279, 2.627227367991459, -1.59963908284846, -0.3122512299247273, -1.0265333654194488, 0.5557085182114885, -0.3206725445321106, -1.4314746067673778, 1.5740113510560039, -0.6566477752702335, 0.31342313477927125, 0.33135361418686027, -1.0489180508346863, -0.2670759024309527, 0.4683952221006179, 0.04918061587657951, 1.239814741442417, 2.2239462179369296, 1.8507671783064434, 1.756319462015174, -0.6577450354719728, 2.2795431083561626, -0.492273906928334, 0.7045614632761499, 0.11260553216111485],), DynamicPPL.DefaultContext()))

We can see which of the variables in a model have been conditioned with DynamicPPL.conditioned:

DynamicPPL.conditioned(partially_conditioned)
(x = [0.8488780584442736, -0.31936138249336765, -1.3982098801744465, -0.05198933163879332, -1.1465116601038348, -0.6306168227545849, 0.6862766694322289, -0.5485073478947856, -0.17212004616875684, 1.2883226251958486  …  0.04918061587657951, 1.239814741442417, 2.2239462179369296, 1.8507671783064434, 1.756319462015174, -0.6577450354719728, 2.2795431083561626, -0.492273906928334, 0.7045614632761499, 0.11260553216111485],)
Note

Sometimes it is helpful to define convenience functions for conditioning on some variable(s). For instance, in this example we might want to define a version of gdemo that conditions on some observations of x:

gdemo(x::AbstractVector{<:Real}) = gdemo(length(x)) | (; x)

For illustrative purposes, however, we do not use this function in the examples below.

Probabilities and Densities

We often want to calculate the (unnormalized) probability density for an event. This probability might be a prior, a likelihood, or a posterior (joint) density. DynamicPPL provides convenient functions for this. To begin, let’s define a model gdemo, condition it on a dataset, and draw a sample. The returned sample only contains μ, since the value of x has already been fixed:

model = gdemo(length(dataset)) | (x=dataset,)

Random.seed!(124)
sample = rand(model)
(μ = -0.6680014719649068,)

We can then calculate the joint probability of a set of samples (here drawn from the prior) with logjoint.

logjoint(model, sample)
-181.7247437162069

For models with many variables rand(model) can be prohibitively slow since it returns a NamedTuple of samples from the prior distribution of the unconditioned variables. We recommend working with samples of type DataStructures.OrderedDict in this case (which Turing re-exports, so can be used directly):

Random.seed!(124)
sample_dict = rand(OrderedDict, model)
OrderedDict{VarName, Any} with 1 entry:
  μ => -0.668001

logjoint can also be used on this sample:

logjoint(model, sample_dict)
-181.7247437162069

The prior probability and the likelihood of a set of samples can be calculated with the functions logprior and loglikelihood respectively. The log joint probability is the sum of these two quantities:

logjoint(model, sample) ≈ loglikelihood(model, sample) + logprior(model, sample)
true
logjoint(model, sample_dict) ≈ loglikelihood(model, sample_dict) + logprior(model, sample_dict)
true

Example: Cross-validation

To give an example of the probability interface in use, we can use it to estimate the performance of our model using cross-validation. In cross-validation, we split the dataset into several equal parts. Then, we choose one of these sets to serve as the validation set. Here, we measure fit using the cross entropy (Bayes loss).1 (For the sake of simplicity, in the following code, we enforce that nfolds must divide the number of data points. For a more competent implementation, see MLUtils.jl.)

# Calculate the train/validation splits across `nfolds` partitions, assume `length(dataset)` divides `nfolds`
function kfolds(dataset::Array{<:Real}, nfolds::Int)
    fold_size, remaining = divrem(length(dataset), nfolds)
    if remaining != 0
        error("The number of folds must divide the number of data points.")
    end
    first_idx = firstindex(dataset)
    last_idx = lastindex(dataset)
    splits = map(0:(nfolds - 1)) do i
        start_idx = first_idx + i * fold_size
        end_idx = start_idx + fold_size
        train_set_indices = [first_idx:(start_idx - 1); end_idx:last_idx]
        return (view(dataset, train_set_indices), view(dataset, start_idx:(end_idx - 1)))
    end
    return splits
end

function cross_val(
    dataset::Vector{<:Real};
    nfolds::Int=5,
    nsamples::Int=1_000,
    rng::Random.AbstractRNG=Random.default_rng(),
)
    # Initialize `loss` in a way such that the loop below does not change its type
    model = gdemo(1) | (x=[first(dataset)],)
    loss = zero(logjoint(model, rand(rng, model)))

    for (train, validation) in kfolds(dataset, nfolds)
        # First, we train the model on the training set, i.e., we obtain samples from the posterior.
        # For normally-distributed data, the posterior can be computed in closed form.
        # For general models, however, typically samples will be generated using MCMC with Turing.
        posterior = Normal(mean(train), 1)
        samples = rand(rng, posterior, nsamples)

        # Evaluation on the validation set.
        validation_model = gdemo(length(validation)) | (x=validation,)
        loss += sum(samples) do sample
            logjoint(validation_model, (μ=sample,))
        end
    end

    return loss
end

cross_val(dataset)
-212760.30282411768
Back to top

Footnotes

  1. See ParetoSmooth.jl for a faster and more accurate implementation of cross-validation than the one provided here.↩︎