Fix #15

workaround to create Multivarite normal distributions from inpositive definite covariance matrix
r9y9 · Oct 19, 2016 · 8acd2ba · 8acd2ba
1 parent 29e8334
commit 8acd2ba
Show file tree

Hide file tree

Showing 3 changed files with 9 additions and 2 deletions.
diff --git a/REQUIRE b/REQUIRE
@@ -3,5 +3,6 @@ Compat
 HDF5 0.4.0
 JLD
 StatsBase
+StatsFuns
 Distributions 0.6.2
 MelGeneralizedCepstrums 0.0.3
diff --git a/src/VoiceConversion.jl b/src/VoiceConversion.jl
@@ -3,6 +3,7 @@
 module VoiceConversion
 
 using StatsBase
+using StatsFuns
 using Distributions
 using MelGeneralizedCepstrums
 using HDF5, JLD

diff --git a/src/gmm.jl b/src/gmm.jl
@@ -3,11 +3,16 @@
 typealias GMM{Cov,Mean} MixtureModel{Multivariate,Continuous,MvNormal{Cov,Mean}}
 
 # proxy to MixtureModel{Multivariate,Continuous,MvNormal{Cov, Mean}}
-function GaussianMixtureModel(means, covars, weights)
+function GaussianMixtureModel(means, covars, weights; forcepd::Bool=true)
     n_components = size(means, 2)
     normals = Array(MvNormal, n_components)
     for m=1:n_components
-        normals[m] = MvNormal(means[:,m], covars[:,:,m])
+        # force covariance matrix to be positive definite to avoid error on
+        # creating PDMat internally
+        # TODO: is this really correct?
+        # caused by https://github.com/JuliaLang/julia/pull/16799
+        covar = forcepd ? Array(Hermitian(covars[:,:,m])) : covars[:,:,m]
+        normals[m] = MvNormal(means[:,m], covar)
     end
     MixtureModel(normals, weights)
 end