Merge branch 'master' into torfjelde/gibbs-new-improv

Author: yebai

.JuliaFormatter.toml

@@ -10,7 +10,7 @@ ignore = [
     "src/mcmc/abstractmcmc.jl",
     "test/experimental/gibbs.jl",
     "test/test_utils/numerical_tests.jl",
-    # https://github.com/TuringLang/Turing.jl/pull/2218/files 
+    # https://github.com/TuringLang/Turing.jl/pull/2218/files
     "src/mcmc/Inference.jl",
     "test/mcmc/Inference.jl",
     # https://github.com/TuringLang/Turing.jl/pull/1887 # Enzyme PR

.github/workflows/Tests.yml

@@ -67,21 +67,43 @@ jobs:
           echo "Julia version: ${{ matrix.version }}"
           echo "Number of threads: ${{ matrix.num_threads }}"
           echo "Test arguments: ${{ matrix.test-args }}"
+      - name: (De)activate coverage analysis
+        run: echo "COVERAGE=${{ matrix.version == '1' && matrix.os == 'ubuntu-latest' && matrix.num_threads == 2 }}" >> "$GITHUB_ENV"
+        shell: bash
       - uses: actions/checkout@v4
       - uses: julia-actions/setup-julia@v2
         with:
           version: '${{ matrix.version }}'
           arch: ${{ matrix.arch }}
-      - uses: actions/cache@v4
-        env:
-          cache-name: cache-artifacts
-        with:
-          path: ~/.julia/artifacts
-          key: ${{ runner.os }}-test-${{ env.cache-name }}-${{ hashFiles('**/Project.toml') }}
-          restore-keys: |
-            ${{ runner.os }}-test-${{ env.cache-name }}-
-            ${{ runner.os }}-test-
-            ${{ runner.os }}-
-      - uses: julia-actions/julia-buildpkg@latest
+      - uses: julia-actions/cache@v1
+      - uses: julia-actions/julia-buildpkg@v1
+      # TODO: Use julia-actions/julia-runtest when test_args are supported
+      # Custom calls of Pkg.test tend to miss features such as e.g. adjustments for CompatHelper PRs
+      # Ref https://github.com/julia-actions/julia-runtest/pull/73  
       - name: Call Pkg.test
-        run: julia --color=yes --depwarn=yes --check-bounds=yes --threads=${{ matrix.num_threads }} --project=@. -e 'import Pkg; Pkg.test(; test_args=ARGS)' -- ${{ matrix.test-args }}
+        run: julia --color=yes --inline=yes --depwarn=yes --check-bounds=yes --threads=${{ matrix.num_threads }} --project=@. -e 'import Pkg; Pkg.test(; coverage=parse(Bool, ENV["COVERAGE"]), test_args=ARGS)' -- ${{ matrix.test-args }}
+      - uses: julia-actions/julia-processcoverage@v1
+        if: ${{ env.COVERAGE }}
+      - uses: codecov/codecov-action@v4
+        if: ${{ env.COVERAGE }}
+        with:
+          fail_ci_if_error: true
+          token: ${{ secrets.CODECOV_TOKEN }}
+          file: lcov.info
+      - uses: coverallsapp/github-action@v2
+        if: ${{ env.COVERAGE }}
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          path-to-lcov: lcov.info
+          flag-name: run-${{ join(matrix.*, '-') }}
+          parallel: true
+
+  finish:
+    needs: test
+    if: ${{ always() }}
+    runs-on: ubuntu-latest
+    steps:
+    - name: Coveralls Finished
+      uses: coverallsapp/github-action@v2
+      with:
+        parallel-finished: true

README.md

@@ -15,13 +15,6 @@ https://turinglang.org/docs/
 
 See [releases](https://github.com/TuringLang/Turing.jl/releases).
 
-## Want to contribute?
-
-Turing was originally created and is now managed by Hong Ge. Current and past Turing team members include [Hong Ge](http://mlg.eng.cam.ac.uk/hong/), [Kai Xu](http://mlg.eng.cam.ac.uk/?portfolio=kai-xu), [Martin Trapp](http://martint.blog), [Mohamed Tarek](https://github.com/mohamed82008), [Cameron Pfiffer](https://business.uoregon.edu/faculty/cameron-pfiffer), [Tor Fjelde](http://retiredparkingguard.com/about.html).
-You can see the complete list on Github: https://github.com/TuringLang/Turing.jl/graphs/contributors.
-
-Turing is an open source project so if you feel you have some relevant skills and are interested in contributing, please get in touch. See the [Contributing](https://turinglang.org/dev/docs/contributing/guide) page for details on the process. You can contribute by opening issues on Github, implementing things yourself, and making a pull request. We would also appreciate example models written using Turing.
-
 ## Issues and Discussions
 
 Issues related to bugs and feature requests are welcome on the [issues page](https://github.com/TuringLang/Turing.jl/issues), while discussions and questions about statistical applications and theory should place on the [Discussions page](https://github.com/TuringLang/Turing.jl/discussions) or [our channel](https://julialang.slack.com/messages/turing/) (`#turing`) in the Julia Slack chat. If you do not have an invitation to Julia's Slack, you can get one by going [here](https://julialang.org/slack/).

src/mcmc/Inference.jl

@@ -240,6 +240,15 @@ DynamicPPL.getlogp(t::Transition) = t.lp
 # Metadata of VarInfo object
 metadata(vi::AbstractVarInfo) = (lp = getlogp(vi),)
 
+# TODO: Implement additional checks for certain samplers, e.g.
+# HMC not supporting discrete parameters.
+function _check_model(model::DynamicPPL.Model)
+    return DynamicPPL.check_model(model; error_on_failure=true)
+end
+function _check_model(model::DynamicPPL.Model, alg::InferenceAlgorithm)
+    return _check_model(model)
+end
+
 #########################################
 # Default definitions for the interface #
 #########################################
@@ -258,8 +267,10 @@ function AbstractMCMC.sample(
     model::AbstractModel,
     alg::InferenceAlgorithm,
     N::Integer;
+    check_model::Bool=true,
     kwargs...
 )
+    check_model && _check_model(model, alg)
     return AbstractMCMC.sample(rng, model, Sampler(alg, model), N; kwargs...)
 end
 
@@ -282,8 +293,10 @@ function AbstractMCMC.sample(
     ensemble::AbstractMCMC.AbstractMCMCEnsemble,
     N::Integer,
     n_chains::Integer;
+    check_model::Bool=true,
     kwargs...
 )
+    check_model && _check_model(model, alg)
     return AbstractMCMC.sample(rng, model, Sampler(alg, model), ensemble, N, n_chains;
                                kwargs...)
 end

test/mcmc/Inference.jl

@@ -559,6 +559,28 @@ using Turing
         @test all(xs[:, 1] .=== [1, missing, 3])
         @test all(xs[:, 2] .=== [missing, 2, 4])
     end
+
+    @testset "check model" begin
+        @model function demo_repeated_varname()
+            x ~ Normal(0, 1)
+            x ~ Normal(x, 1)
+        end
+
+        @test_throws ErrorException sample(
+            demo_repeated_varname(), NUTS(), 1000; check_model=true
+        )
+        # Make sure that disabling the check also works.
+        @test (sample(
+            demo_repeated_varname(), Prior(), 10; check_model=false
+        ); true)
+
+        @model function demo_incorrect_missing(y)
+            y[1:1] ~ MvNormal(zeros(1), 1)
+        end
+        @test_throws ErrorException sample(
+            demo_incorrect_missing([missing]), NUTS(), 1000; check_model=true
+        )
+    end
 end
 
 end

test/mcmc/gibbs.jl

@@ -50,7 +50,9 @@ using Turing.RandomMeasures: ChineseRestaurantProcess, DirichletProcess
         Random.seed!(100)
         alg = Gibbs(CSMC(15, :s), HMC(0.2, 4, :m; adtype=adbackend))
         chain = sample(gdemo(1.5, 2.0), alg, 10_000)
-        check_numerical(chain, [:s, :m], [49 / 24, 7 / 6]; atol=0.15)
+        check_numerical(chain, [:m], [7 / 6]; atol=0.15)
+        # Be more relaxed with the tolerance of the variance.
+        check_numerical(chain, [:s], [49 / 24]; atol=0.35)
 
         Random.seed!(100)
 

test/mcmc/hmc.jl

@@ -319,7 +319,7 @@ using Turing
 
         # The discrepancies in the chains are in the tails, so we can't just compare the mean, etc.
         # KS will compare the empirical CDFs, which seems like a reasonable thing to do here.
-        @test pvalue(ApproximateTwoSampleKSTest(vec(results), vec(results_prior))) > 0.01
+        @test pvalue(ApproximateTwoSampleKSTest(vec(results), vec(results_prior))) > 0.001
     end
 end
 

test/mcmc/is.jl

@@ -46,7 +46,7 @@ using Turing
         ref = reference(n)
 
         Random.seed!(seed)
-        chain = sample(model, alg, n)
+        chain = sample(model, alg, n; check_model=false)
         sampled = get(chain, [:a, :b, :lp])
 
         @test vec(sampled.a) == ref.as

test/mcmc/mh.jl

@@ -44,30 +44,41 @@ GKernel(var) = (x) -> Normal(x, sqrt.(var))
         # c6 = sample(gdemo_default, s6, N)
     end
     @testset "mh inference" begin
+        # Set the initial parameters, because if we get unlucky with the initial state,
+        # these chains are too short to converge to reasonable numbers.
+        discard_initial = 1000
+        initial_params = [1.0, 1.0]
+
         Random.seed!(125)
         alg = MH()
-        chain = sample(gdemo_default, alg, 10_000)
+        chain = sample(gdemo_default, alg, 10_000; discard_initial, initial_params)
         check_gdemo(chain; atol=0.1)
 
         Random.seed!(125)
         # MH with Gaussian proposal
         alg = MH((:s, InverseGamma(2, 3)), (:m, GKernel(1.0)))
-        chain = sample(gdemo_default, alg, 10_000)
+        chain = sample(gdemo_default, alg, 10_000; discard_initial, initial_params)
         check_gdemo(chain; atol=0.1)
 
         Random.seed!(125)
         # MH within Gibbs
         alg = Gibbs(MH(:m), MH(:s))
-        chain = sample(gdemo_default, alg, 10_000)
+        chain = sample(gdemo_default, alg, 10_000; discard_initial, initial_params)
         check_gdemo(chain; atol=0.1)
 
         Random.seed!(125)
         # MoGtest
         gibbs = Gibbs(
             CSMC(15, :z1, :z2, :z3, :z4), MH((:mu1, GKernel(1)), (:mu2, GKernel(1)))
         )
-        chain = sample(MoGtest_default, gibbs, 500)
-        check_MoGtest_default(chain; atol=0.15)
+        chain = sample(
+            MoGtest_default,
+            gibbs,
+            500;
+            discard_initial=100,
+            initial_params=[1.0, 1.0, 0.0, 0.0, 1.0, 4.0],
+        )
+        check_MoGtest_default(chain; atol=0.2)
     end
 
     # Test MH shape passing.