Fix 173

[yebai]

No description provided.

[xukai92]

sample.jl (i.e. Gibbs with HMC + PG on gdemo) works on Julia 0.6 with 2 changes:

1. Make model wrapper call and sample function call separate (ffb9407);
2. Update to ForwardDiff.jl's new signature (2a9466d);

1. is just a work around for the test. Need to fix it.

[xukai92]

I figure out how to resolve the issue by using Base.invokelatest() - but we may need to make change to the signature of model because Base.invokelatest() doesn't support keyword arguments. (Note: out model signature is like model(vi=vi, sampler=spl) which is keyword argument based.)

[xukai92]

• Fix 1. by changing the model signature and using invokelatest().
• Fix warning messages as many as possible

[xukai92]

Most of the test are passed now. But Julia 0.6 (or other changes?) may lead to some minor changes to some cases. E.g.

p[i,j] ~ dist

generates a VarName called p[1, 2] (with space) now, which was p[1,2] (without space). This means we need to changes many of the tests/benchmarks for results fetching.

Any idea why?

[yebai]

This might be because we introduced a mechanism for translating ~ operator into a macro in Julia 0.6.

[xukai92]

@yebai I see.

All the tests passed on my local Linux machine now. I changed some deprecated functions but now all. I will do that in the next few days.

[yebai]

I pumped Julia version to 0.6 on Travis and Appveyor. The tests run on my Mac as well, but seems quite slow at the moment.

Ps. I got an error on Mac

[runtests.jl] "ad.jl/ad1.jl" is running
[Turing]:  Assume - `s` is a parameter
 @~(::ANY, ::ANY) at compiler.jl:76
[Turing]:  Assume - `m` is a parameter
 @~(::ANY, ::ANY) at compiler.jl:76
ERROR: LoadError: LoadError: MethodError: no method matching ##ad_test_model#1140()
Closest candidates are:
  ##ad_test_model#1140(::Turing.VarInfo, ::Union{Turing.Sampler, Void}) at /Users/yebai/.julia/v0.6/Turing/test/ad.jl/ad1.jl:10
Stacktrace:
 [1] include_from_node1(::String) at ./loading.jl:569
 [2] include(::String) at ./sysimg.jl:14
 [3] macro expansion at /Users/yebai/.julia/v0.6/Turing/test/runtests.jl:86 [inlined]
 [4] anonymous at ./<missing>:?
 [5] include_from_node1(::String) at ./loading.jl:569
 [6] include(::String) at ./sysimg.jl:14
 [7] process_options(::Base.JLOptions) at ./client.jl:305
 [8] _start() at ./client.jl:371
while loading /Users/yebai/.julia/v0.6/Turing/test/ad.jl/ad1.jl, in expression starting on line 20
while loading /Users/yebai/.julia/v0.6/Turing/test/runtests.jl, in expression starting on line 74

[xukai92]

Yes this is a bug in the test. My last commit fixes the test.

This is becauseI didn't actually implement the callback functions (I tried that naively but the compiler doesn't work as I expect) but always use modelf(vi::VarInfo, sampler::Sampler). In empty argument cases before I changed those codes to modelf(VarInfo(), nothing). It's wired my Linux doesn't give the error.

• implement callback functions

[xukai92]

Known unfixed warnings:

• WARNING: consume is now deprecated. Use Channels for inter-task communication.

[xukai92]

I also noticed Turing.jl becomes very slow after migrated to 0.6. I was guessing that's because those deprecations inside model function (which will be called mutiple times during model evatluation), e.g. exp vs exp.

julia> a = randn(100)

julia> @time for _ = 1:1000 exp(a) end
  0.496221 seconds (83.00 k allocations: 12.161 MiB, 1.33% gc time)

julia> @time for _ = 1:1000 exp.(a) end
  0.088814 seconds (49.79 k allocations: 3.175 MiB)

I've fixed most of them - seems become a bit faster.

[xukai92]

@yebai I finally resolve the callback issue.

In terms of the deprecation, I guess the only one unsolved is the Channel one. I don't completely understand what's going on with it. Do you think we can simply replace all Task with Channel?

[yebai]

In terms of the deprecation, I guess the only one unsolved is the Channel one. I don't completely understand what's going on with it. Do you think we can simply replace all Task with Channel?

I'm not sure about this yet; will take a look at the new Task interface.

[yebai]

Yes, we can use Channel in Julia 0.6, but we don't need to drop Task. Basically, the consume/produce API is replaced by take!/put!.

f(_c::Channel{Any}) = begin
      for i =1:1000
      println(i);put!(_c,i);
      end
end

a = Channel(f)

take!(a)

The part related to task copying might be a bit tricky. I will look into this further later.

Ref: JuliaLang/julia#19841

[yebai]

@xukai92 Seems the tests are failing. Could you take a look at Travis?

[yebai]

@xukai92 I still haven't figured how to use channels correctly when tasks are copied. In the meantime, I implemented a temporary hack so that we can continue to use produce before I have a better solution.

[xukai92]

List of un-fixed warnings (taken from https://travis-ci.org/yebai/Turing.jl/jobs/301057755):

• floor
• tanh
• abs
• logpdf
• log

[xukai92]

@yebai Plz have a look at this change - do you agree what we had before was a bug?

[xukai92]

@yebai This is actually (mathematically) correct ... as there is another bug - we didn't use customised init here for vectorised assume at all. I will push another commit to fix it.

[yebai]

yes - it looks correct to me.