Minor bug fixes

src/qda.jl

@@ -40,7 +40,7 @@ function qda!{T<:BlasReal,U<:Integer}(
         γ::Nullable{T}
     )
     H = center_classes!(X, M, y)
-    isnull(λ) ? class_whiteners!(H, y, k, γ, get(λ)) : class_whiteners!(H, y, k, γ)
+    isnull(λ) ? class_whiteners!(H, y, γ) : class_whiteners!(H, y, γ, get(λ))
 end
 
 doc"`qda(X, y; M, lambda, gamma, priors)` Fits a regularized quadratic discriminant model to the 
@@ -83,7 +83,6 @@ function discriminants_qda{T<:BlasReal}(
         end
     end
     δ
-
 end
 
 doc"`discriminants(Model, Z)` Uses `Model` on input `Z` to product the class discriminants."
@@ -95,74 +94,3 @@ doc"`classify(Model, Z)` Uses `Model` on input `Z`."
 function classify{T<:BlasReal}(mod::ModelQDA{T}, Z::Matrix{T})
     mapslices(indmax, discriminants(mod, Z), 2)
 end
-
-# Create an array of class scatter matrices
-#   H is centered data matrix (with respect to class means)
-#   y is one-based vector of class IDs
-#=
-function class_covariances{T<:BlasReal,U<:Integer}(H::Matrix{T}, y::Vector{U}, 
-                                                   n_k::Vector{Int64} = class_counts(y))
-    k = length(n_k)
-    Σ_k = Array(Array{T,2}, k)  # Σ_k[i] = H_i'H_i/(n_i-1)
-    for i = 1:k
-        Σ_k[i] = BLAS.syrk!('U', 'T', one(T)/(n_k[i]-1), H[y .== i,:], zero(T), Array(T,p,p))
-        symml!(Σ_k[i])
-    end
-    Σ_k
-end
-=#
-
-
-
-
-# Use eigendecomposition to generate class whitening transform
-#   Σ_k is array of references to each Σ_i covariance matrix
-#   λ is regularization parameter in [0,1]. λ = 0 is no regularization.
-#=
-function class_whiteners!{T<:BlasReal}(Σ_k::Vector{Matrix{T}}, γ::T)
-    for i = 1:length(Σ_k)
-        tol = eps(T) * prod(size(Σ_k[i])) * maximum(Σ_k[i])
-        Λ_i, V_i = LAPACK.syev!('V', 'U', Σ_k[i])  # Overwrite Σ_k with V such that VΛVᵀ = Σ_k
-        if γ > 0
-            λ_avg = mean(Λ_i)  # Shrink towards average eigenvalue
-            for j = 1:length(Λ_i)
-                Λ_i[j] = (1-γ)*Λ_i[j] + γ*λ_avg  # Σ = VΛVᵀ => (1-γ)Σ + γI = V((1-γ)Λ + γI)Vᵀ
-            end
-        end
-        all(Λ_i .>= tol) || error("Rank deficiency detected in class $i with tolerance $tol.")
-        scale!(V_i, one(T) ./ sqrt(Λ_i))  # Scale V so it whitens H*V where H is centered X
-    end
-    Σ_k
-end
-=#
-
-# Fit regularized quadratic discriminant model. Returns whitening matrices for all classes.
-#   X in uncentered data matrix
-#   M is matrix of class means (one per row)
-#   y is one-based vector of class IDs
-#   λ is regularization parameter in [0,1]. λ = 0 is no regularization. See documentation.
-#   γ is shrinkage parameter in [0,1]. γ = 0 is no shrinkage. See documentation.
-#=
-function qda!{T<:BlasReal,U<:Integer}(X::Matrix{T}, M::Matrix{T}, y::Vector{U}, λ::T, γ::T)
-    k    = maximum(y)
-    n_k  = class_counts(y, k)
-    n, p = size(X)
-    H    = center_classes!(X, M, y)
-    w_σ  = 1 ./ vec(sqrt(var(H, 1)))  # Variance normalizing factor for columns of H
-    scale!(H, w_σ)
-    Σ_k  = class_covariances(H, y, n_k)
-    if λ > 0
-        Σ = scale!(H'H, one(T)/(n-1))
-        for i = 1:k 
-            regularize!(Σ_k[i], λ, Σ)
-        end
-    end
-    W_k = class_whiteners!(Σ_k, γ)
-    for i = 1:k
-        scale!(w_σ, W_k[i])  # scale rows of W_k
-    end
-    W_k
-end
-=#
-
-

test/test_qda.jl

@@ -46,3 +46,18 @@ for T in FloatingPointTypes
         end
     end
 end
+
+info("Testing ", MOD.qda!)
+for T in FloatingPointTypes
+    X_tmp = copy(convert(Matrix{T}, X))
+    M_tmp = convert(Matrix{T}, M)
+    H_tmp = convert(Matrix{T}, H)
+
+    W_k = MOD.qda!(copy(X_tmp), copy(M_tmp), y, Nullable{T}(), Nullable{T}())
+    W_k_tmp = MOD.class_whiteners!(copy(H_tmp), y, Nullable{T}())
+
+    for i in eachindex(W_k)
+        @test_approx_eq W_k[i] W_k_tmp[i]
+    end
+
+end