Merge pull request #35 from JuliaAI/dev

For a 0.3 release

Author: ablaom

.github/workflows/ci.yml

@@ -17,7 +17,7 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - '1.0'
+          - '1.6'
           - '1'
         os:
           - ubuntu-latest

Project.toml

@@ -1,7 +1,7 @@
 name = "MLJMultivariateStatsInterface"
 uuid = "1b6a4a23-ba22-4f51-9698-8599985d3728"
 authors = ["Anthony D. Blaom <[email protected]>", "Thibaut Lienart <[email protected]>", "Okon Samuel <[email protected]>"]
-version = "0.2.2"
+version = "0.3.0"
 
 [deps]
 Distances = "b4f34e82-e78d-54a5-968a-f98e89d6e8f7"
@@ -13,9 +13,9 @@ StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
 [compat]
 Distances = "^0.9,^0.10"
 MLJModelInterface = "^0.3.5,^0.4, 1.0"
-MultivariateStats = "0.7, 0.8"
+MultivariateStats = "0.9"
 StatsBase = "0.32, 0.33"
-julia = "1"
+julia = "1.6"
 
 [extras]
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"

src/models/decomposition_models.jl

@@ -9,14 +9,14 @@ $PCA_DESCR
 
 # Keyword Parameters
 
-- `maxoutdim::Int=0`: maximum number of output dimensions, uses the smallest dimension of 
+- `maxoutdim::Int=0`: maximum number of output dimensions, uses the smallest dimension of
     training feature matrix if 0 (default).
-- `method::Symbol=:auto`: method to use to solve the problem, one of `:auto`,`:cov` 
+- `method::Symbol=:auto`: method to use to solve the problem, one of `:auto`,`:cov`
     or `:svd`
 - `pratio::Float64=0.99`: ratio of variance preserved
-- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: if set to nothing(default)  
-    centering will be computed and applied, if set to `0` no 
-    centering(assumed pre-centered), if a vector is passed, the centering is done with 
+- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: if set to nothing(default)
+    centering will be computed and applied, if set to `0` no
+    centering(assumed pre-centered), if a vector is passed, the centering is done with
     that vector.
 """
 @mlj_model mutable struct PCA <: MMI.Unsupervised
@@ -34,21 +34,20 @@ function MMI.fit(model::PCA, verbosity::Int, X)
     Xarray = MMI.matrix(X)
     mindim = minimum(size(Xarray))
     maxoutdim = model.maxoutdim == 0 ? mindim : model.maxoutdim
-    # NOTE: copy/transpose
     fitresult = MS.fit(
-        MS.PCA, transpose(Xarray);
+        MS.PCA, Xarray';
         method=model.method,
         pratio=model.pratio,
         maxoutdim=maxoutdim,
         mean=model.mean
     )
     cache = nothing
     report = (
-        indim=MS.indim(fitresult),
-        outdim=MS.outdim(fitresult),
+        indim=MS.size(fitresult,1),
+        outdim=MS.size(fitresult,2),
         tprincipalvar=MS.tprincipalvar(fitresult),
         tresidualvar=MS.tresidualvar(fitresult),
-        tvar=MS.tvar(fitresult),
+        tvar=MS.var(fitresult),
         mean=copy(MS.mean(fitresult)),
         principalvars=copy(MS.principalvars(fitresult))
     )
@@ -74,14 +73,14 @@ $KPCA_DESCR
 
 # Keyword Parameters
 
-- `maxoutdim::Int = 0`: maximum number of output dimensions, uses the smallest 
+- `maxoutdim::Int = 0`: maximum number of output dimensions, uses the smallest
     dimension of training feature matrix if 0 (default).
-- `kernel::Function=(x,y)->x'y`: kernel function of 2 vector arguments x and y, returns a 
+- `kernel::Function=(x,y)->x'y`: kernel function of 2 vector arguments x and y, returns a
     scalar value
-- `solver::Symbol=:auto`: solver to use for the eigenvalues, one of `:eig`(default), 
+- `solver::Symbol=:auto`: solver to use for the eigenvalues, one of `:eig`(default),
     `:eigs`
-- `inverse::Bool=false`: perform calculation for inverse transform
-- `beta::Real=1.0`: strength of the ridge regression that learns the inverse transform 
+- `inverse::Bool=true`: perform calculations needed for inverse transform
+- `beta::Real=1.0`: strength of the ridge regression that learns the inverse transform
     when inverse is true
 - `tol::Real=0.0`: Convergence tolerance for eigs solver
 - `maxiter::Int=300`: maximum number of iterations for eigs solver
@@ -90,7 +89,7 @@ $KPCA_DESCR
     maxoutdim::Int = 0::(_ ≥ 0)
     kernel::Union{Nothing, Function} = default_kernel
     solver::Symbol = :eig::(_ in (:eig, :eigs))
-    inverse::Bool = false
+    inverse::Bool = true
     beta::Real = 1.0::(_ ≥ 0.0)
     tol::Real = 1e-6::(_ ≥ 0.0)
     maxiter::Int = 300::(_ ≥ 1)
@@ -102,7 +101,7 @@ function MMI.fit(model::KernelPCA, verbosity::Int, X)
     # default max out dim if not given
     maxoutdim = model.maxoutdim == 0 ? mindim : model.maxoutdim
     fitresult = MS.fit(
-        MS.KernelPCA, 
+        MS.KernelPCA,
         permutedims(Xarray);
         kernel=model.kernel,
         maxoutdim=maxoutdim,
@@ -113,9 +112,9 @@ function MMI.fit(model::KernelPCA, verbosity::Int, X)
     )
     cache  = nothing
     report = (
-        indim=MS.indim(fitresult),
-        outdim=MS.outdim(fitresult),
-        principalvars=copy(MS.principalvars(fitresult))
+        indim=MS.size(fitresult,1),
+        outdim=MS.size(fitresult,2),
+        principalvars=copy(MS.eigvals(fitresult))
     )
     return fitresult, cache, report
 end
@@ -143,17 +142,16 @@ $ICA_DESCR
 
 - `k::Int=0`: number of independent components to recover, set automatically if `0`
 - `alg::Symbol=:fastica`: algorithm to use (only `:fastica` is supported at the moment)
-- `fun::Symbol=:tanh`: approximate neg-entropy functor, via the function 
-    `MultivariateStats.icagfun`, one of `:tanh` and `:gaus`
+- `fun::Symbol=:tanh`: approximate neg-entropy function, one of `:tanh`, `:gaus`
 - `do_whiten::Bool=true`: whether to perform pre-whitening
 - `maxiter::Int=100`: maximum number of iterations
 - `tol::Real=1e-6`: convergence tolerance for change in matrix W
-- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: mean to use, if nothing (default) 
-    centering is computed andapplied, if zero, no centering, a vector of means can 
+- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: mean to use, if nothing (default)
+    centering is computed andapplied, if zero, no centering, a vector of means can
     be passed
-- `winit::Union{Nothing,Matrix{<:Real}}=nothing`: initial guess for matrix `W` either 
-    an empty matrix (random initilization of `W`), a matrix of size `k × k` (if `do_whiten` 
-    is true), a matrix of size `m × k` otherwise. If unspecified i.e `nothing` an empty 
+- `winit::Union{Nothing,Matrix{<:Real}}=nothing`: initial guess for matrix `W` either
+    an empty matrix (random initilization of `W`), a matrix of size `k × k` (if `do_whiten`
+    is true), a matrix of size `m × k` otherwise. If unspecified i.e `nothing` an empty
     `Matrix{<:Real}` is used.
 """
 @mlj_model mutable struct ICA <: MMI.Unsupervised
@@ -168,14 +166,19 @@ $ICA_DESCR
 end
 
 function MMI.fit(model::ICA, verbosity::Int, X)
+    icagfun(fname::Symbol, ::Type{T} = Float64) where T<:Real=
+    fname == :tanh ? MS.Tanh{T}(1.0) :
+    fname == :gaus ? MS.Gaus{T}() :
+    error("Unknown gfun $(fname)")
+
     Xarray = MMI.matrix(X)
     n, p = size(Xarray)
     m = min(n, p)
     k = ifelse(model.k ≤ m, model.k, m)
     fitresult = MS.fit(
-        MS.ICA, transpose(Xarray), k;
+        MS.ICA, Xarray', k;
         alg=model.alg,
-        fun=MS.icagfun(model.fun, eltype(Xarray)),
+        fun=icagfun(model.fun, eltype(Xarray)),
         do_whiten=model.do_whiten,
         maxiter=model.maxiter,
         tol=model.tol,
@@ -184,8 +187,8 @@ function MMI.fit(model::ICA, verbosity::Int, X)
     )
     cache = nothing
     report = (
-        indim=MS.indim(fitresult),
-        outdim=MS.outdim(fitresult),
+        indim=MS.size(fitresult,1),
+        outdim=MS.size(fitresult,2),
         mean=copy(MS.mean(fitresult))
     )
     return fitresult, cache, report
@@ -211,14 +214,14 @@ $PPCA_DESCR
 
 # Keyword Parameters
 
-- `maxoutdim::Int=0`: maximum number of output dimensions, uses max(no_of_features - 1, 1) 
+- `maxoutdim::Int=0`: maximum number of output dimensions, uses max(no_of_features - 1, 1)
     if 0 (default).
 - `method::Symbol=:ml`: method to use to solve the problem, one of `:ml`, `:em`, `:bayes`.
 - `maxiter::Int=1000`: maximum number of iterations.
 - `tol::Real=1e-6`: convergence tolerance.
-- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: if set to nothing(default)  
-    centering will be computed and applied, if set to `0` no 
-    centering(assumed pre-centered), if a vector is passed, the centering is done with 
+- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: if set to nothing(default)
+    centering will be computed and applied, if set to `0` no
+    centering(assumed pre-centered), if a vector is passed, the centering is done with
     that vector.
 """
 @mlj_model mutable struct PPCA <: MMI.Unsupervised
@@ -233,9 +236,8 @@ function MMI.fit(model::PPCA, verbosity::Int, X)
     Xarray = MMI.matrix(X)
     def_dim = max(1, size(Xarray, 2) - 1)
     maxoutdim = model.maxoutdim == 0 ? def_dim : model.maxoutdim
-    # NOTE: copy/transpose
     fitresult = MS.fit(
-        MS.PPCA, transpose(Xarray);
+        MS.PPCA, Xarray';
         method=model.method,
         tol=model.tol,
         maxiter=model.maxiter,
@@ -244,8 +246,8 @@ function MMI.fit(model::PPCA, verbosity::Int, X)
     )
     cache = nothing
     report = (
-        indim=MS.indim(fitresult),
-        outdim=MS.outdim(fitresult),
+        indim=MS.size(fitresult,1),
+        outdim=MS.size(fitresult,2),
         tvar=MS.var(fitresult),
         mean=copy(MS.mean(fitresult)),
         loadings=MS.loadings(fitresult)
@@ -273,14 +275,14 @@ $PPCA_DESCR
 # Keyword Parameters
 
 - `method::Symbol=:cm`: Method to use to solve the problem, one of `:ml`, `:em`, `:bayes`.
-- `maxoutdim::Int=0`: Maximum number of output dimensions, uses max(no_of_features - 1, 1) 
+- `maxoutdim::Int=0`: Maximum number of output dimensions, uses max(no_of_features - 1, 1)
     if 0 (default).
 - `maxiter::Int=1000`: Maximum number of iterations.
 - `tol::Real=1e-6`: Convergence tolerance.
 - `eta::Real=tol`: Variance lower bound
-- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: If set to nothing(default)  
-    centering will be computed and applied, if set to `0` no 
-    centering(assumed pre-centered), if a vector is passed, the centering is done with 
+- `mean::Union{Nothing, Real, Vector{Float64}}=nothing`: If set to nothing(default)
+    centering will be computed and applied, if set to `0` no
+    centering(assumed pre-centered), if a vector is passed, the centering is done with
     that vector.
 """
 @mlj_model mutable struct FactorAnalysis <: MMI.Unsupervised
@@ -296,9 +298,8 @@ function MMI.fit(model::FactorAnalysis, verbosity::Int, X)
     Xarray = MMI.matrix(X)
     def_dim = max(1, size(Xarray, 2) - 1)
     maxoutdim = model.maxoutdim == 0 ? def_dim : model.maxoutdim
-    # NOTE: copy/transpose
     fitresult = MS.fit(
-        MS.FactorAnalysis, transpose(Xarray);
+        MS.FactorAnalysis, Xarray';
         method=model.method,
         maxiter=model.maxiter,
         tol=model.tol,
@@ -308,8 +309,8 @@ function MMI.fit(model::FactorAnalysis, verbosity::Int, X)
     )
     cache = nothing
     report = (
-        indim=MS.indim(fitresult),
-        outdim=MS.outdim(fitresult),
+        indim=MS.size(fitresult,1),
+        outdim=MS.size(fitresult,2),
         variance=MS.var(fitresult),
         covariance_matrix=MS.cov(fitresult),
         mean=MS.mean(fitresult),
@@ -344,17 +345,17 @@ for (M, MFitResultType) in model_types
     end
 
     @eval function MMI.transform(::$M, fr::$MFitResultType, X)
-        # X is n x d, need to transpose twice
+        # X is n x d, need to take adjoint twice
         Xarray = MMI.matrix(X)
-        Xnew = transpose(MS.transform(fr, transpose(Xarray)))
+        Xnew = MS.predict(fr, Xarray')'
         return MMI.table(Xnew, prototype=X)
     end
 
     if hasmethod(MS.reconstruct, Tuple{MFitResultType{Float64}, Matrix{Float64}})
         @eval function MMI.inverse_transform(::$M, fr::$MFitResultType, Y)
-            # X is n x p, need to transpose twice
+            # X is n x p, need to take adjoint twice
             Yarray = MMI.matrix(Y)
-            Ynew = transpose(MS.reconstruct(fr, transpose(Yarray)))
+            Ynew = MS.reconstruct(fr, Yarray')'
             return MMI.table(Ynew, prototype=Y)
         end
     end