Last active
November 18, 2023 22:26
-
-
Save torfjelde/b484c689db7bf8d62d3a07ea6260827e to your computer and use it in GitHub Desktop.
Using`loo` in Turing.jl
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
julia> using ArviZ, Turing | |
julia> J = 8; | |
julia> y = [28.0, 8.0, -3.0, 7.0, -1.0, 1.0, 18.0, 12.0]; | |
julia> σ = [15.0, 10.0, 16.0, 11.0, 9.0, 11.0, 10.0, 18.0]; | |
julia> schools = [ | |
"Choate", | |
"Deerfield", | |
"Phillips Andover", | |
"Phillips Exeter", | |
"Hotchkiss", | |
"Lawrenceville", | |
"St. Paul's", | |
"Mt. Hermon" | |
]; | |
julia> nwarmup, nsamples, nchains = 1000, 1000, 4; | |
julia> @model function model(J, y, σ, ::Type{TV} = Vector{Float64}) where {TV} | |
μ ~ Normal(0, 5) | |
τ ~ truncated(Cauchy(0, 5), 0, Inf) | |
θ = TV(undef, J) | |
θ .~ Normal(μ, τ) | |
# If we want the POINT-WISE likelihoods, we need to refer by different symbols, | |
# i.e. cannot use `MvNormal`. | |
for j = 1:J | |
y[j] ~ Normal(θ[j], σ[j]) | |
end | |
end; | |
julia> param_mod = model(J, y, σ); | |
julia> sampler = NUTS(nwarmup, 0.8); | |
julia> turing_chns = sample( | |
param_mod, | |
sampler, | |
nwarmup + nsamples; | |
progress = false, | |
); | |
┌ Info: Found initial step size | |
└ ϵ = 0.4 | |
julia> # Extract the loglikelihoods, and convert into form which ArviZ.jl accepts | |
ℓ = DynamicPPL.elementwise_loglikelihoods(param_mod, turing_chns) | |
Dict{DynamicPPL.VarName,Array{Float64,1}} with 8 entries: | |
y[3] => [-3.72618, -3.7039, -3.70514, -3.84801, -3.95597, -3.73676, -3.70973, -3.70654, -3.69412, -3.78229 … -3.73184, -3.72254, -3.97129, -3.81096, -3.73516, -3.7398, -3.77288, -3.7099, -3.69716, -3.69155] | |
y[1] => [-5.16847, -5.37776, -5.33898, -4.71719, -4.4846, -5.20381, -5.38348, -5.52579, -5.52824, -4.96575 … -4.82861, -4.82866, -4.65692, -4.99942, -4.61389, -5.07198, -5.26182, -5.36327, -5.89517, -6.03677] | |
y[4] => [-3.51839, -3.50616, -3.52167, -3.32892, -3.33368, -3.3899, -3.46509, -3.62935, -3.56666, -3.34411 … -3.31696, -3.31689, -3.64576, -3.49002, -3.31885, -3.31966, -3.37919, -3.69016, -3.47888, -3.57988] | |
y[5] => [-3.13272, -3.19451, -3.14482, -3.3349, -3.76755, -3.18067, -3.12868, -3.11639, -3.11626, -3.26104 … -3.16744, -3.16209, -3.42108, -3.11728, -3.22843, -3.41312, -3.20184, -3.2123, -3.13486, -3.39949] | |
y[6] => [-3.32513, -3.31684, -3.31948, -3.40729, -3.5989, -3.32907, -3.31691, -3.35456, -3.33646, -3.35431 … -3.33286, -3.3366, -3.49421, -3.3826, -3.36658, -3.32905, -3.50545, -3.42483, -3.47718, -3.44098] | |
y[7] => [-4.53587, -4.90501, -5.01885, -3.89998, -3.68522, -4.53326, -4.84988, -5.1454, -4.98813, -4.195 … -4.00933, -3.9955, -3.98242, -3.91085, -4.58724, -4.00131, -5.01847, -4.40975, -4.14685, -3.79092] | |
y[8] => [-4.01703, -4.03218, -4.01876, -3.87603, -3.8295, -3.95203, -4.01166, -4.07744, -4.04852, -3.89826 … -3.9914, -4.0061, -3.81249, -3.82682, -4.09209, -3.86704, -4.13824, -3.96275, -4.17168, -4.22254] | |
y[2] => [-3.46757, -3.50013, -3.45111, -3.24411, -3.23219, -3.35582, -3.45173, -3.70297, -3.59883, -3.31984 … -3.256, -3.27523, -3.25873, -3.33843, -3.23371, -3.34246, -3.30696, -3.87759, -3.40094, -3.53952] | |
julia> ℓ_mat = reduce(hcat, values(ℓ)); | |
julia> ℓ_arr = reshape(ℓ_mat, 1, size(ℓ_mat)...); # (chain_idx, sample_idx, parameter_idx) | |
julia> idata = ArviZ.from_mcmcchains( | |
turing_chns, | |
coords = Dict("school" => schools), | |
dims = Dict( | |
"y" => ["school"], | |
"σ" => ["school"], | |
"θ" => ["school"], | |
), | |
library = "Turing", | |
log_likelihood = Dict("y" => ℓ_arr) | |
) | |
InferenceData with groups: | |
> posterior | |
> log_likelihood | |
> sample_stats | |
julia> ArviZ.loo(idata) | |
1×7 DataFrame | |
│ Row │ loo │ loo_se │ p_loo │ n_samples │ n_data_points │ warning │ loo_scale │ | |
│ │ Float64 │ Float64 │ Float64 │ Int64 │ Int64 │ Bool │ String │ | |
├─────┼─────────┼─────────┼──────────┼───────────┼───────────────┼─────────┼───────────┤ | |
│ 1 │ -30.717 │ 1.38963 │ 0.781578 │ 1000 │ 8 │ 0 │ log │ |
Hey! Glad to hear:)
So I actually ended up adding this to ArviZ.jl documentation (we'll include examples of this in Turing tutorials soon too) here: https://arviz-devs.github.io/ArviZ.jl/stable/quickstart/#Additional-information-in-Turing.jl. That example uses multiple chains so should hopefully provide you with an answer!
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hey @torfjelde, this is super useful for me actually! The only I wonder, is what should it look like when you have multiple chains (because of using MCMCThreads for example). So far I figure it should maybe be something like:
I am sure there is a cleaner more robust way of writing what I am trying to achieve anyway, but it seems like you would want it to end up with: n_chains x n_samples x n_parameters as the dimensions of the ℓ_arr variable I believe right?