More tests for the Gibbs sampler

[turinglang]

More tests for the Gibbs sampler

[xukai92]

Good suggestion! Will do it along with bug fixing.

[arnim]

Hi,

great library!
While I'm rather new to it I stumbled upon a small issue

@model gauss1 begin
    N = length(x)
    lam = ones(Float64,N)
    mu ~ Normal(0, sqrt(1000));
    for i = 1:N
        lam[i] ~ Gamma(.001, .001)
        x[i] ~ Normal(mu, sqrt(1 / (sqrt(lam[i]))))
    end
    mu
end

x = [-27.020, 3.570, 8.191, 9.898, 9.603, 9.945, 10.056]
chain = sample(gauss1,  SMC(300))

work. However

@model gauss2 begin
    N = length(x)
    lam = ones(Float64,N)
    mu ~ Normal(0, sqrt(1000));
    
    for i = 1:N
        lam[i] ~ Gamma(.001, .001)
        x[i] ~ Normal(mu, sqrt(1 / (sqrt(lam[i]))))
    end
    mu, lam
end

chain = sample(gauss2,  SMC(300))

gives me

`
MethodError: no method matching mcmcstats(::Array{Array{Float64,1},1})
Closest candidates are:
mcmcstats{T<:Real}(::Array{T<:Real,1}) at /home/juser/.julia/v0.5/Turing/src/core/io.jl:160
mcmcstats{T<:Real}(::Array{T<:Real,2}) at /home/juser/.julia/v0.5/Turing/src/core/io.jl:168

in show(::IOContext{Base.AbstractIOBuffer{Array{UInt8,1}}}, ::Turing.Chain) at /home/juser/.julia/v0.5/Turing/src/core/io.jl:85
in limitstringmime(::MIME{Symbol("text/plain")}, ::Turing.Chain) at /opt/julia_packages/.julia/v0.5/IJulia/src/inline.jl:25
in display_dict(::Turing.Chain) at /opt/julia_packages/.julia/v0.5/IJulia/src/execute_request.jl:28
in execute_request(::ZMQ.Socket, ::IJulia.Msg) at /opt/julia_packages/.julia/v0.5/IJulia/src/execute_request.jl:188
in eventloop(::ZMQ.Socket) at /opt/julia_packages/.julia/v0.5/IJulia/src/eventloop.jl:8
in (::IJulia.##13#19)() at ./task.jl:360
`

What would you recommend?
Im at the current head at a6d4cc8

Best
Arnim

[yebai]

Hi Arnim,

there is a small bug in the print function for Chain, which stores samples from a sampling algorithm. The following code with an extra semi-colon runs fine

chain = sample(gauss2,  SMC(300)); 

We will fix this bug in the next release.

Thanks

[yebai]

BTW, the current master branch contains some (major) ongoing refactoring of the interface, the result should not be trusted too much for the moment. This should gets better in the next new release.

[arnim]

Thanks @yebai for your quick response.
For now something like in stan-models/naive_bayes.jl:25 should do the trick of inspecting.

Let me know when I can help.
Really like the intuitive way of defining programs.

Best
Arnim

[xukai92]

@yebai For more examples in the future, we can have separate files for data, model and inference, which would also make tests for front-end/interface only easy. I adapted my previous 3 Stan models following this. Please check https://github.com/xukai92/Turing-exps/tree/master/stan-models.

[yebai]

@xukai92 Thanks Kai. I have added some models myself. I suggest that we keep these example models under the example-models folder in the main repo for the moment. It will be easier for us to add and run these models within the main repo.

[xukai92]

@yebai Agreed. I've made the update to the main repo too.

[xukai92]

Goal for this isuee: make Naive Bayes and LDA work

[xukai92]

Closed by PR #162