Turing.jl: errors during testing (HMC?)

[mkarikom]

Platform: Windows 10, Julia 1.1.0

running the following produces an error:
add Turing
test Turing

.... lines about things being updated

[bbc10e6e] + DynamicHMC v1.0.2+ #master (https://github.com/tpapp/DynamicHMC.jl)
[HMC] Finished with
Running time = 1.179192493000001;
Accept rate = 1.0;
#lf / sample = 0.9966666666666667;
#evals / sample = 2.9966666666666666;
pre-cond. metric = [1.0].
[HMC] Finished with
Running time = 1.9699e-5;
Accept rate = 1.0;
#lf / sample = 0.0;
#evals / sample = 2.0;
pre-cond. metric = [1.0].
[HMC] Finished with
Running time = 0.6601277;
Accept rate = 1.0;
#lf / sample = 0.5;
#evals / sample = 2.5;
pre-cond. metric = [1.0].
[-9.42436, -7.7163]
[-9.42436, -7.7163]
[ Info: [PG] Finished with
[ Info: Running time = 0.3503446859999997;
ERROR: Package Turing errored during testing

Using OSX 10.14, Julia 1.1.0, we have a similar result:

[bbc10e6e] + DynamicHMC v1.0.2+ #master (https://github.com/tpapp/DynamicHMC.jl)
[HMC] Finished with
Running time = 1.6609471790000014;
Accept rate = 0.9966666666666667;
#lf / sample = 0.9966666666666667;
#evals / sample = 2.9966666666666666;
pre-cond. metric = [1.0].
[HMC] Finished with
Running time = 3.3658e-5;
Accept rate = 1.0;
#lf / sample = 0.0;
#evals / sample = 2.0;
pre-cond. metric = [1.0].
[HMC] Finished with
Running time = 0.997483639;
Accept rate = 1.0;
#lf / sample = 0.5;
#evals / sample = 2.5;
pre-cond. metric = [1.0].
[0.200685, 1.2323]
[0.200685, 1.2323]
[ Info: [PG] Finished with
[ Info: Running time = 0.5600100159999993;
signal (11): Segmentation fault: 11
in expression starting at /Users/au/.julia/packages/Turing/FTRCE/test/compiler.jl/beta_binomial.jl:26
jl_swap_fiber at /Users/osx/buildbot/slave/package_osx64/build/src/task.c:778
Allocations: 140303015 (Pool: 140281852; Big: 21163); GC: 317
ERROR: Package Turing errored during testing

[mohamed82008]

Turing doesn't support Julia 1.1 yet, see https://github.com/TuringLang/Turing.jl/issues/657. Please try it on Julia 1.0.3.

[mkarikom]

Turing doesn't support Julia 1.1 yet, see #657. Please try it on Julia 1.0.3.

Ok, thanks!

[yebai]

We've made some progress; a fix should be coming soon.

[mohamed82008]

This should be fixed in the latest release.

[ManojGopalkrishnan]

I'm on Julia 1.4.1 and Windows 10, I got the same failure message "Package Turing errored during testing." The first error is

"Simplex Tracker, Zygote and ReverseDiff (with and without caching) AD: Error During Test at ~.julia\packages\Turing\cReBm\test\core\ad.jl:269
Got exception outside of a @test
MethodError: no method matching AxisArrays.AxisArray(::Array{Float64,3}; iter=1:1:1000, var=["theta[2]", "theta[4]", "theta[3]", "theta[1]", "lp", "n_steps", "is_accept", "acceptance_rate", "log_density", "hamiltonian_energy", "hamiltonian_energy_error", "step_size", "nom_step_size"], chain=1:1)"

and then it goes on like that for a few pages.

[devmotion]

This seems to be the same error as https://github.com/TuringLang/Turing.jl/issues/1220. Can you check that you use the latest version of MCMCChains (i.e., MCMCChains 3.0.12)?

[ManojGopalkrishnan]

Thank you I updated MCMCChains but it's still showing only 3.0.11 . How do I get 3.0.12?

[devmotion]

Maybe some other package is holding it back. By running

julia> import Pkg; Pkg.add("[email protected]")

Pkg should tell you which other package conflicts with 3.0.12.

[ManojGopalkrishnan]

Ok figured this out.
Pkg.add(Pkg.PackageSpec(name="MCMCChains", version="3.0.12"))
gave output

Resolving package versions...
Installed IntervalSets ─ v0.4.0
Installed MCMCChains ─── v3.0.12
Installed AxisArrays ─── v0.4.2
Updating ~\.julia\environments\v1.4\Project.toml
[39de3d68] ↑ AxisArrays v0.4.0 ⇒ v0.4.2
[c7f686f2] ↑ MCMCChains v3.0.11 ⇒ v3.0.12
Updating ~\.julia\environments\v1.4\Manifest.toml
[39de3d68] ↑ AxisArrays v0.4.0 ⇒ v0.4.2
[8197267c] ↓ IntervalSets v0.5.0 ⇒ v0.4.0
[c7f686f2] ↑ MCMCChains v3.0.11 ⇒ v3.0.12

IntervalSets seems to be the conflicting package here. Also, this fixed my problems with Turing, thanks!

[devmotion]

Great that it works now!

[ManojGopalkrishnan]

Umm still having trouble :-( I tried this code

using Turing
using StatsPlots

# Define a simple Normal model with unknown mean and variance.
@model gdemo(x, y) = begin
  s ~ InverseGamma(2, 3)
  m ~ Normal(0, sqrt(s))
  x ~ Normal(m, sqrt(s))
  y ~ Normal(m, sqrt(s))
end

#  Run sampler, collect results
chn = sample(gdemo(1.5, 2), HMC(0.1, 5), 1000)

# Summarise results (currently requires the master branch from MCMCChains)
describe(chn)

# Plot and save results
p = plot(chn)
savefig("gdemo-plot.png")

and got the same error:

MethodError: no method matching AxisArray(::Array{Float64,3}; iter=1:1:1000, var=["m", "s", "lp", "n_steps", "is_accept", "acceptance_rate", "log_density", "hamiltonian_energy", "hamiltonian_energy_error", "step_size", "nom_step_size"], chain=1:1)

MCMCChains is on v3.0.12, and Turing passed the Pkg.test("Turing") checks okay.

[devmotion]

Hmm are you sure you are using the latest version of MCMCChains? Such models are tested in the Turing tests, so I'm a bit surprised that they pass and your example fails :open_mouth:

[ManojGopalkrishnan]

I am surprised as well. Yes, I just updated MCMCChains, and it shows version 3.0.12 and AxisArrays is v0.4.2 .

[ManojGopalkrishnan]

Ok it's working now! So Pkg.test("Turing") was working (tests passed) after updating MCMCChains, but I was getting error messages on simple examples. Apparently I had to shut down the IJulia notebook, and restart it. All very strange.

[devmotion]

Oh, yes, after updating packages that are already loaded you should restart Julia (or in your case the IJulia kernel). Glad you figured it out!

[devmotion]

Running Pkg.test("Turing") starts a different Julia process, so that's why the updated package was used there.