support 32-bit MulticlassLDA (#34)

src/lda.jl

@@ -104,7 +104,7 @@ function multiclass_lda_stats{T<:AbstractFloat}(nc::Int, X::DenseMatrix{T}, y::A
     Sw = A_mul_Bt(Z, Z)
 
     # compute between-class scattering
-    mean = cmeans * (cweights ./ n)
+    mean = cmeans * (cweights ./ T(n))
     U = scale!(cmeans .- mean, sqrt.(cweights))
     Sb = A_mul_Bt(U, U)
 
@@ -137,7 +137,7 @@ transform{T<:AbstractFloat}(M::MulticlassLDA, x::AbstractVecOrMat{T}) = M.proj'x
 function fit{T<:AbstractFloat}(::Type{MulticlassLDA}, nc::Int, X::DenseMatrix{T}, y::AbstractVector{Int};
              method::Symbol=:gevd,
              outdim::Int=min(size(X,1), nc-1),
-             regcoef::T=1.0e-6)
+             regcoef::T=T(1.0e-6))
 
     multiclass_lda(multiclass_lda_stats(nc, X, y);
                    method=method,
@@ -148,7 +148,7 @@ end
 function multiclass_lda{T<:AbstractFloat}(S::MulticlassLDAStats{T};
                         method::Symbol=:gevd,
                         outdim::Int=min(size(X,1), S.nclasses-1),
-                        regcoef::T=1.0e-6)
+                        regcoef::T=T(1.0e-6))
 
     P = mclda_solve(S.Sb, S.Sw, method, outdim, regcoef)
     MulticlassLDA(P, P'S.cmeans, S)

test/mclda.jl

@@ -111,19 +111,21 @@ test_approx_eq_vecs(P1, P2)
 
 ## LDA
 
-M = fit(MulticlassLDA, nc, X, y; method=:gevd, regcoef=lambda)
-@test indim(M) == d
-@test outdim(M) == nc - 1
-@test projection(M) ≈ P1
-@test M.pmeans ≈ M.proj'cmeans
-@test transform(M, X) ≈ M.proj'X
-
-M = fit(MulticlassLDA, nc, X, y; method=:whiten, regcoef=lambda)
-@test indim(M) == d
-@test outdim(M) == nc - 1
-# @test projection(M) P2  # signs may change
-@test M.pmeans ≈ M.proj'cmeans
-@test transform(M, X) ≈ M.proj'X
+for T in (Float32, Float64)
+    M = fit(MulticlassLDA, nc, convert(Matrix{T}, X), y; method=:gevd, regcoef=convert(T, lambda))
+    @test indim(M) == d
+    @test outdim(M) == nc - 1
+    @test projection(M) ≈ P1
+    @test M.pmeans ≈ M.proj'cmeans
+    @test transform(M, X) ≈ M.proj'X
+
+    M = fit(MulticlassLDA, nc, convert(Matrix{T}, X), y; method=:whiten, regcoef=convert(T, lambda))
+    @test indim(M) == d
+    @test outdim(M) == nc - 1
+    # @test projection(M) P2  # signs may change
+    @test M.pmeans ≈ M.proj'cmeans
+    @test transform(M, X) ≈ M.proj'X
+end
 
 
 ## High-dimensional LDA (subspace LDA)