Description
Currently from_mcmcchains assumes that all variable names are single-bracket-delimited or dot-delimited. However, quite complicated names are possible:
julia> using Turing, LinearAlgebra
julia> @model function foo()
a ~ Normal()
bar = (b=Matrix{typeof(a)}(undef, 2, 3), c=Vector{typeof(a)}(undef, 3))
bar.b[1, :] .~ Normal(a)
bar.b[2, 1:2] ~ MvNormal(I(2))
bar.b[2, 3:end] .~ Normal(a)
bar.c[end:-1:1] .~ Normal(a)
end
foo (generic function with 2 methods)
julia> chns = sample(foo(), NUTS(), 1_000)
┌ Info: Found initial step size
└ ϵ = 1.6
Sampling 100%|█████████████████████████████████████████████████████████████████████████| Time: 0:00:01
Chains MCMC chain (1000×22×1 Array{Float64, 3}):
Iterations = 501:1:1500
Number of chains = 1
Samples per chain = 1000
Wall duration = 1.21 seconds
Compute duration = 1.21 seconds
parameters = a, bar.b[1,:][1], bar.b[1,:][2], bar.b[1,:][3], bar.b[2,1:2][1], bar.b[2,1:2][2], bar.b[2,3:3][1], bar.c[3:-1:1][1], bar.c[3:-1:1][2], bar.c[3:-1:1][3]
internals = lp, n_steps, is_accept, acceptance_rate, log_density, hamiltonian_energy, hamiltonian_energy_error, max_hamiltonian_energy_error, tree_depth, numerical_error, step_size, nom_step_size
Summary Statistics
parameters mean std naive_se mcse ess rhat ess_per_sec
Symbol Float64 Float64 Float64 Float64 Float64 Float64 Float64
a 0.0278 0.9966 0.0315 0.0724 139.5528 1.0152 115.1426
bar.b[1,:][1] 0.0178 1.4143 0.0447 0.0841 205.0380 1.0096 169.1733
bar.b[1,:][2] 0.0268 1.3678 0.0433 0.0765 292.3916 1.0072 241.2472
bar.b[1,:][3] 0.0395 1.3866 0.0438 0.0852 234.2846 1.0061 193.3041
bar.b[2,1:2][1] 0.0395 1.0091 0.0319 0.0260 1017.5267 0.9990 839.5435
bar.b[2,1:2][2] -0.0270 0.9839 0.0311 0.0260 1495.1411 0.9990 1233.6147
bar.b[2,3:3][1] 0.0506 1.3798 0.0436 0.0724 347.4877 1.0032 286.7061
bar.c[3:-1:1][1] 0.0156 1.4357 0.0454 0.0835 275.8889 1.0079 227.6311
bar.c[3:-1:1][2] -0.0010 1.4186 0.0449 0.0720 362.6178 1.0072 299.1896
bar.c[3:-1:1][3] 0.0146 1.4061 0.0445 0.0783 333.0312 1.0069 274.7783
Quantiles
parameters 2.5% 25.0% 50.0% 75.0% 97.5%
Symbol Float64 Float64 Float64 Float64 Float64
a -1.8465 -0.6502 -0.0164 0.7037 2.0664
bar.b[1,:][1] -2.7699 -0.9711 0.0028 0.9771 2.7813
bar.b[1,:][2] -2.5305 -0.9451 0.0408 0.9232 2.7366
bar.b[1,:][3] -2.6779 -0.9212 0.0311 1.0146 2.7213
bar.b[2,1:2][1] -1.8999 -0.6486 0.0397 0.7054 2.1453
bar.b[2,1:2][2] -1.9572 -0.7017 -0.0031 0.6454 1.8877
bar.b[2,3:3][1] -2.6416 -0.8997 -0.0014 0.9453 3.0213
bar.c[3:-1:1][1] -2.8470 -0.9299 -0.0128 0.9509 2.8362
bar.c[3:-1:1][2] -2.8268 -0.9619 0.0671 0.9812 2.7441
bar.c[3:-1:1][3] -2.7313 -0.9267 0.0001 0.9474 2.8913If we call from_mcmcchains on this, we get an uninformative error.
Ideally we would like to get an InferenceData with bar.b and bar.c as variables. However, since the modeler can arbitrarily index and reindex and call getproperty to make arbitrarily complicated types, always doing the right thing is probably not possible. Also, MCMCChains's own machinery for combining flattened parameters into parameter arrays doesn't do a great job here.
For the short term then, I think it makes the most sense to raise an informative error of splitting by brackets produces anything more complicated than a tuple of integer indices. If users find this constraining and open issues, we can discuss supporting slightly more complicated indexing syntaxes.