Merge pull request #143 from JuliaAI/dev

For a 0.9.1 release

Project.toml

@@ -1,7 +1,7 @@
 name = "MLJLinearModels"
 uuid = "6ee0df7b-362f-4a72-a706-9e79364fb692"
 authors = ["Thibaut Lienart <tlienart@me.com>"]
-version = "0.9.0"
+version = "0.9.1"
 
 [deps]
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"
@@ -16,7 +16,7 @@ Parameters = "d96e819e-fc66-5662-9728-84c9c7592b0a"
 DocStringExtensions = "0.8, 0.9"
 IterativeSolvers = "0.8, 0.9"
 LinearMaps = "2.6, 3.2"
-MLJModelInterface = "0.3, 0.4, 1.0"
+MLJModelInterface = "1.4"
 Optim = "0.20, 0.21, 1"
 Parameters = "0.12"
 julia = "1.6, 1"

src/mlj/classifiers.jl

@@ -1,33 +1,58 @@
 #= ===================
    LOGISTIC CLASSIFIER
-   =================== =#
+=================== =#
 
 """
-Logistic Classifier (typically called "Logistic Regression"). This model is
-a standard classifier for both binary and multiclass classification.
-In the binary case it corresponds to the LogisticLoss, in the multiclass to the
-Multinomial (softmax) loss. An elastic net penalty can be applied with
-overall objective function
+$(doc_header(LogisticClassifier))
 
-``L(y, Xθ) + n⋅λ|θ|₂²/2 + n⋅γ|θ|₁``
+This model is more commonly known as "logistic regression". It is a standard classifier
+for both binary and multiclass classification.  The objective function applies either a
+logistic loss (binary target) or multinomial (softmax) loss, and has a mixed L1/L2
+penalty:
 
-where ``L`` is either the logistic or multinomial loss and ``λ`` and ``γ`` indicate
-the strength of the L2 (resp. L1) regularisation components and
-``n`` is the number of samples `size(X, 1)`.
-With `scale_penalty_with_samples = false` the objective function is
-``L(y, Xθ) + λ|θ|₂²/2 + γ|θ|₁``
+``L(y, Xθ) + n⋅λ|θ|₂²/2 + n⋅γ|θ|₁``.
 
-## Parameters
+Here ``L`` is either `MLJLinearModels.LogisticLoss` or `MLJLinearModels.MultiClassLoss`,
+``λ`` and ``γ`` indicate
+the strength of the L2 (resp. L1) regularization components and
+``n`` is the number of training observations.
+
+With `scale_penalty_with_samples = false` the objective function is instead
+
+``L(y, Xθ) + λ|θ|₂²/2 + γ|θ|₁``.
+
+# Training data
+
+In MLJ or MLJBase, bind an instance `model` to data with
+
+    mach = machine(model, X, y)
+
+where:
+
+- `X` is any table of input features (eg, a `DataFrame`) whose columns
+  have `Continuous` scitype; check column scitypes with `schema(X)`
+
+- `y` is the target, which can be any `AbstractVector` whose element
+  scitype is `<:OrderedFactor` or `<:Multiclass`; check the scitype
+  with `scitype(y)`
+
+Train the machine using `fit!(mach, rows=...)`.
+
+
+# Hyperparameters
 
 $TYPEDFIELDS
 
 $(example_docstring("LogisticClassifier", nclasses = 2))
+
+See also [`MultinomialClassifier`](@ref).
+
 """
 @with_kw_noshow mutable struct LogisticClassifier <: MMI.Probabilistic
-    "strength of the regulariser if `penalty` is `:l2` or `:l1` and strength of the L2
-    regulariser if `penalty` is `:en`."
+    "strength of the regularizer if `penalty` is `:l2` or `:l1` and strength of the L2
+    regularizer if `penalty` is `:en`."
     lambda::Real             = eps()
-    "strength of the L1 regulariser if `penalty` is `:en`."
+    "strength of the L1 regularizer if `penalty` is `:en`."
     gamma::Real              = 0.0
     "the penalty to use, either `:l2`, `:l1`, `:en` (elastic net) or `:none`."
     penalty::SymStr          = :l2
@@ -37,7 +62,19 @@ $(example_docstring("LogisticClassifier", nclasses = 2))
     penalize_intercept::Bool = false
     "whether to scale the penalty with the number of samples."
     scale_penalty_with_samples::Bool = true
-    "type of solver to use, default if `nothing`."
+    """some instance of `MLJLinearModels.S` where `S` is one of: `LBFGS`, `Newton`,
+    `NewtonCG`, `ProxGrad`; but subject to the following restrictions:
+
+    - If `penalty = :l2`, `ProxGrad` is disallowed. Otherwise, `ProxyGrad` is the only
+      option.
+
+    - Unless `scitype(y) <: Finite{2}` (binary target) `Newton` is disallowed.
+
+    If `solver = nothing` (default) then `ProxGrad(accel=true)` (FISTA) is used,
+    unless `gamma = 0`, in which case `LBFGS()` is used.
+
+    Solver aliases: `FISTA(; kwargs...) = ProxGrad(accel=true, kwargs...)`,
+    `ISTA(; kwargs...) = ProxGrad(accel=false, kwargs...)`"""
     solver::Option{Solver}   = nothing
 end
 
@@ -50,27 +87,49 @@ glr(m::LogisticClassifier, nclasses::Integer) =
                        scale_penalty_with_samples=m.scale_penalty_with_samples,
                        nclasses=nclasses)
 
-descr(::Type{LogisticClassifier}) = "Classifier corresponding to the loss function ``L(y, Xθ) + λ|θ|₂²/2 + γ|θ|₁`` where `L` is the logistic loss."
-
 #= ======================
    MULTINOMIAL CLASSIFIER
    ====================== =#
 
 """
-See `LogisticClassifier`, it's the same except that multiple classes are assumed
-by default. The other parameters are the same.
+$(doc_header(MultinomialClassifier))
+
+This model coincides with [`LogisticClassifier`](@ref), except certain optimizations
+possible in the special binary case will not be applied. Its hyperparameters are
+identical.
+
+# Training data
+
+In MLJ or MLJBase, bind an instance `model` to data with
+
+    mach = machine(model, X, y)
+
+where:
+
+- `X` is any table of input features (eg, a `DataFrame`) whose columns
+  have `Continuous` scitype; check column scitypes with `schema(X)`
+
+- `y` is the target, which can be any `AbstractVector` whose element
+  scitype is `<:OrderedFactor` or `<:Multiclass`; check the scitype
+  with `scitype(y)`
+
+Train the machine using `fit!(mach, rows=...)`.
 
-## Parameters
+
+# Hyperparameters
 
 $TYPEDFIELDS
 
-$(example_docstring("LogisticClassifier", nclasses = 3))
+$(example_docstring("MultinomialClassifier", nclasses = 3))
+
+See also [`LogisticClassifier`](@ref).
+
 """
 @with_kw_noshow mutable struct MultinomialClassifier <: MMI.Probabilistic
-    "strength of the regulariser if `penalty` is `:l2` or `:l1`.
-    Strength of the L2 regulariser if `penalty` is `:en`."
+    "strength of the regularizer if `penalty` is `:l2` or `:l1`.
+    Strength of the L2 regularizer if `penalty` is `:en`."
     lambda::Real             = eps()
-    "strength of the L1 regulariser if `penalty` is `:en`."
+    "strength of the L1 regularizer if `penalty` is `:en`."
     gamma::Real              = 0.0
     "the penalty to use, either `:l2`, `:l1`, `:en` (elastic net) or `:none`."
     penalty::SymStr          = :l2
@@ -80,7 +139,19 @@ $(example_docstring("LogisticClassifier", nclasses = 3))
     penalize_intercept::Bool = false
     "whether to scale the penalty with the number of samples."
     scale_penalty_with_samples::Bool = true
-    "type of solver to use, default if `nothing`."
+    """some instance of `MLJLinearModels.S` where `S` is one of: `LBFGS`,
+    `NewtonCG`, `ProxGrad`; but subject to the following restrictions:
+
+    - If `penalty = :l2`, `ProxGrad` is disallowed. Otherwise, `ProxyGrad` is the only
+      option.
+
+    - Unless `scitype(y) <: Finite{2}` (binary target) `Newton` is disallowed.
+
+    If `solver = nothing` (default) then `ProxGrad(accel=true)` (FISTA) is used,
+    unless `gamma = 0`, in which case `LBFGS()` is used.
+
+    Solver aliases: `FISTA(; kwargs...) = ProxGrad(accel=true, kwargs...)`,
+    `ISTA(; kwargs...) = ProxGrad(accel=false, kwargs...)`"""
     solver::Option{Solver}   = nothing
 end
 
@@ -91,7 +162,3 @@ glr(m::MultinomialClassifier, nclasses::Integer) =
                           penalize_intercept=m.penalize_intercept,
                           scale_penalty_with_samples=m.scale_penalty_with_samples,
                           nclasses=nclasses)
-
-descr(::Type{MultinomialClassifier}) =
-    "Classifier corresponding to the loss function " *
-    "``L(y, Xθ) + λ|θ|₂²/2 + γ|θ|₁`` where `L` is the multinomial loss."

src/mlj/doc_tools.jl

@@ -0,0 +1,40 @@
+const DOC_SOLVERS =
+    "Different solver options exist, as indicated under "*
+    "\"Hyperparameters\" below. "
+
+function example_docstring(m; nclasses = nothing)
+"""
+## Example
+
+    using MLJ
+    X, y = $(nclasses == nothing ? "make_regression()" : "make_blobs(centers = $nclasses)")
+    mach = fit!(machine($m(), X, y))
+    predict(mach, X)
+    fitted_params(mach)
+
+"""
+end
+
+function doc_header(ModelType)
+    name = MLJModelInterface.name(ModelType)
+    human_name = MLJModelInterface.human_name(ModelType)
+
+    """
+        $name
+
+    A model type for constructing a $human_name, based on
+    [MLJLinearModels.jl](https://github.com/alan-turing-institute/MLJLinearModels.jl), and
+    implementing the MLJ model interface.
+
+    From MLJ, the type can be imported using
+
+        $name = @load $name pkg=MLJLinearModels
+
+    Do `model = $name()` to construct an instance with default
+    hyper-parameters.
+
+    """
+end
+
+const DOC_PROXGRAD = "Aliases `ISTA` and `FSTA` correspond to "*
+    "`ProxGrad` with the option `acceleration=false` or `true` respectively. "

src/mlj/interface.jl

@@ -4,19 +4,7 @@ export LinearRegressor, RidgeRegressor, LassoRegressor, ElasticNetRegressor,
 
 const SymStr = Union{Symbol,String}
 
-function example_docstring(m; nclasses = nothing)
-"""
-## Example
-
-    using MLJ
-    X, y = $(nclasses == nothing ? "make_regression()" : "make_blobs(centers = $nclasses)")
-    mach = fit!(machine($m(), X, y))
-    predict(mach, X)
-    fitted_params(mach)
-
-"""
-end
-
+include("doc_tools.jl")
 include("regressors.jl")
 include("classifiers.jl")
 
@@ -121,8 +109,6 @@ function MMI.fitted_params(m::Union{CLF_MODELS...}, (θ, features, classes, c))
     return _fitted_params(θ, features, nothing)
 end
 
-@static VERSION < v"1.1" && (eachrow(A::AbstractVecOrMat) = (view(A, i, :) for i in axes(A, 1)))
-
 coef_vec(W::AbstractMatrix, features) =
     [feature => coef for (feature, coef) in zip(features, eachrow(W))]
 coef_vec(θ::AbstractVector, features) =
@@ -135,30 +121,30 @@ coef_vec(θ::AbstractVector, ::Nothing) = θ
    ======================= =#
 
 MMI.metadata_pkg.(ALL_MODELS,
-    name="MLJLinearModels",
-    uuid="6ee0df7b-362f-4a72-a706-9e79364fb692",
-    url="https://github.com/alan-turing-institute/MLJLinearModels.jl",
-    julia=true,
-    license="MIT",
+    package_name="MLJLinearModels",
+    package_uuid="6ee0df7b-362f-4a72-a706-9e79364fb692",
+    package_url="https://github.com/alan-turing-institute/MLJLinearModels.jl",
+    is_pure_julia=true,
+    package_license="MIT",
     is_wrapper=false)
 
-descr_(M) = descr(M) *
-    "\n→ based on [MLJLinearModels](https://github.com/alan-turing-institute/MLJLinearModels.jl)" *
-    "\n→ do `@load $(MMI.name(M)) pkg=\"MLJLinearModels\" to use the model.`" *
-    "\n→ do `?$(MMI.name(M))` for documentation."
 lp_(M) = "MLJLinearModels.$(MMI.name(M))"
 
 for M in REG_MODELS
-    MMI.metadata_model(M,
-        input=MMI.Table(MMI.Continuous),
-        target=AbstractVector{MMI.Continuous},
-        weights=false,
-        descr=descr_(M), path=lp_(M))
+    MMI.metadata_model(
+        M,
+        input_scitype=MMI.Table(MMI.Continuous),
+        target_scitype=AbstractVector{MMI.Continuous},
+        load_path=lp_(M),
+    )
 end
 for M in CLF_MODELS
-    MMI.metadata_model(M,
-        input=MMI.Table(MMI.Continuous),
-        target=AbstractVector{<:MMI.Finite},
-        weights=false,
-        descr=descr_(M), path=lp_(M))
+    MMI.metadata_model(
+        M,
+        input_scitype=MMI.Table(MMI.Continuous),
+        target_scitype=AbstractVector{<:MMI.Finite},
+        load_path=lp_(M),
+    )
 end
+
+MMI.human_name(::Type{<:LADRegressor}) = "least absolute deviation regressor"