Merge pull request #32 from bcbi/dpa/loglikelihood

Add the `loglikelihood` function

Author: DilumAluthge

.gitignore

@@ -2,8 +2,14 @@
 *.jl.cov
 *.jl.mem
 .DS_Store
+
 /Manifest.toml
+
 /dev/
+
+/docs/Manifest.toml
 /docs/build/
 /docs/site/
-/docs/Manifest.toml
+/docs/src/examples/
+
+/test/Manifest.toml

Project.toml

@@ -1,7 +1,7 @@
 name = "MaximumLikelihoodProblems"
 uuid = "e149cee4-a9c8-4a03-a7b6-d91720412337"
 authors = ["Dilum Aluthge"]
-version = "0.1.3"
+version = "0.1.4"
 
 [deps]
 LogDensityProblems = "6fdf6af0-433a-55f7-b3ed-c6c6e0b8df7c"

docs/make.jl

@@ -5,34 +5,38 @@ import Literate
 import Random
 
 _this_filename = @__FILE__
-_docs_directory = dirname(_this_filename)
-root_directory = dirname(_docs_directory)
-docs_directory = joinpath(root_directory, "docs")
-docs_source_directory = joinpath(root_directory, "docs", "src")
-examples_directory = joinpath(root_directory, "examples")
+_docs = dirname(_this_filename)
+root = dirname(_docs)
+docs = joinpath(root, "docs")
+docs_src = joinpath(root, "docs", "src")
+docs_src_examples = joinpath(root, "docs", "src", "examples")
+examples_directory = joinpath(root, "examples")
 
 examples_list = ["linear_regression",
                  "logistic_regression",
                  "multinomial_logistic_regression"]
 
 for example in examples_list
-    example_filename = joinpath(examples_directory, "$(example).jl")
+    example_filename = joinpath(examples_directory,
+                                "$(example).jl")
     Literate.markdown(example_filename,
-                      docs_source_directory)
+                      docs_src_examples)
 end
 
 Random.seed!(123)
 
+generated_examples = [joinpath("examples", "$(example).md") for example in examples_list]
+
 Documenter.makedocs(;
-    root=docs_directory,
-    modules=[MaximumLikelihoodProblems],
-    format=Documenter.HTML(),
+    root = docs,
+    modules = [MaximumLikelihoodProblems],
+    format = Documenter.HTML(),
     pages = vcat(["Home" => "index.md",],
-                  ["$(example).md" for example in examples_list],
+                  generated_examples,
                   ["api_public.md",]),
-    repo="https://github.com/bcbi/MaximumLikelihoodProblems.jl/blob/{commit}{path}#L{line}",
-    sitename="MaximumLikelihoodProblems.jl",
-    authors="Dilum P. Aluthge",
+    repo = "https://github.com/bcbi/MaximumLikelihoodProblems.jl/blob/{commit}{path}#L{line}",
+    sitename = "MaximumLikelihoodProblems.jl",
+    authors = "Dilum P. Aluthge",
 )
 
 Documenter.deploydocs(;

docs/src/api_public.md

@@ -1,5 +1,10 @@
 # Public API
 
+```@index
+Pages   = ["api_public.md"]
+```
+
 ```@docs
 MaximumLikelihoodProblems.fit
+MaximumLikelihoodProblems.loglikelihood
 ```

examples/linear_regression.jl

@@ -75,3 +75,7 @@ transformed_gradient_problem = LogDensityProblems.ADgradient(:ForwardDiff,
 # σ_hat:
 
 σ_hat = θ_hat[:σ]
+
+# Value of the log likelihood function evaluated at θ_hat:
+
+MaximumLikelihoodProblems.loglikelihood(transformed_gradient_problem, θ_hat)

examples/logistic_regression.jl

@@ -54,3 +54,7 @@ transformed_gradient_problem = LogDensityProblems.ADgradient(:ForwardDiff,
 # β_hat:
 
 β_hat = θ_hat[:β]
+
+# Value of the log likelihood function evaluated at θ_hat:
+
+MaximumLikelihoodProblems.loglikelihood(transformed_gradient_problem, θ_hat)

examples/multinomial_logistic_regression.jl

@@ -70,3 +70,7 @@ transformed_gradient_problem = LogDensityProblems.ADgradient(:ForwardDiff,
 
 num_covariates = size(β_hat, 1)
 β_hat_with_base_class = hcat(zeros(num_covariates), β_hat)
+
+# Value of the log likelihood function evaluated at θ_hat:
+
+MaximumLikelihoodProblems.loglikelihood(transformed_gradient_problem, θ_hat)

src/MaximumLikelihoodProblems.jl

@@ -5,5 +5,6 @@ include("public.jl")
 include("types.jl")
 
 include("fit.jl")
+include("loglikelihood.jl")
 
 end # module

src/fit.jl

@@ -11,6 +11,8 @@ const default_tolerance = 1e-10
 """
     fit(transformed_gradient_problem, theta_hat_initial; kwargs...)
 
+Find the maximum likelihood estimatator for the parameters `theta`.
+
 # Arguments
 - `transformed_gradient_problem`
 - `theta_hat_initial`
@@ -62,14 +64,14 @@ function _fit(transformed_gradient_problem,
     end
     for iter = 1:max_iterations
         theta_hat_old = theta_hat_new
-        loglikelihood, gradient = LogDensityProblems.logdensity_and_gradient(transformed_gradient_problem,
-                                                                             theta_hat_old)
+        log_likelihood_value, gradient = LogDensityProblems.logdensity_and_gradient(transformed_gradient_problem,
+                                                                                    theta_hat_old)
         update = learning_rate * gradient
         theta_hat_new = theta_hat_old + update
         update_norm = sum(abs, update)
         if show_progress_meter
             showvalues = [(:iteration, iter),
-                          (:loglikelihood, loglikelihood)]
+                          (:loglikelihood, log_likelihood_value)]
             ProgressMeter.update!(progress,
                                   update_norm;
                                   showvalues = showvalues)

src/loglikelihood.jl

@@ -0,0 +1,29 @@
+import LogDensityProblems
+import TransformVariables
+
+"""
+    loglikelihood(transformed_gradient_problem, theta_hat)
+
+Return the value of the log likelihood function evaluated at `theta_hat`.
+
+# Arguments
+- `transformed_gradient_problem`
+- `theta_hat`
+"""
+function loglikelihood(transformed_gradient_problem,
+                       theta)
+    transformed_log_density = parent(transformed_gradient_problem)
+    transformation = transformed_log_density.transformation
+    theta_inversetransformed = TransformVariables.inverse(transformation,
+                                                          theta)
+    result = _loglikelihood(transformed_gradient_problem,
+                            theta_inversetransformed)
+    return result
+end
+
+function _loglikelihood(transformed_gradient_problem,
+                        theta)
+    log_likelihood_value, gradient = LogDensityProblems.logdensity_and_gradient(transformed_gradient_problem,
+                                                                                theta)
+    return log_likelihood_value
+end