search

davidavdav / GaussianMixtures.jl

Large scale Gaussian Mixture Models
Other
99 stars 38 forks source link

Returning the same Gaussians #64

Closed giadasp closed 4 years ago

giadasp commented 4 years ago

Dear David,

I'm trying to fit a bivariate Gaussian mixture with 3 components to my data. The issue is that the resulting components are all equal (same mean vector and var/covar matrix), also the three weights are equal. I've initiliazed the gmm with gmm=GMM(3,2;kind=:full) and then I trained it by em!(gmm,X,nIter=10) I saw in the history that is converging. my X is 2000x2. Am I doing anything wrong? Thank you in advance for your support.

Giada

davidavdav commented 4 years ago

I think you're missing initialization of the Gaussians, your constructor initializes with 0's and identity covariance, and em! is not going to be able to separate these identical mixtures. So usually you initialize the components using something like k-means.

The trick here is to use the training constructor:

gmm = GMM(3, X, nIter=10)

which does the k-means, followed by EM for you.

giadasp commented 4 years ago

Thank you for the prompt reply, I have this error with nIter=10

ERROR: MethodError: no method matching GMM(::Array{Float64,2}, ::Int64; kind=:full, nIter=10)
Closest candidates are:
  GMM(::Int64, ::Int64; kind) at C:\Users\...\.julia\packages\GaussianMixtures\3jRIL\src\gmms.jl:12 got unsupported keyword argument "nIter"

and this one without

ERROR: MethodError: no method matching GMM(::Array{Float64,2}, ::Int64; kind=:full)
Closest candidates are:
  GMM(::Int64, ::Int64; kind) at C:\Users\...\.julia\packages\GaussianMixtures\3jRIL\src\gmms.jl:12
  GMM(::Union{Array{T,2}, Data{T,VT} where VT<:Union{AbstractString, Array{T,2} where T}}; kind) where T<:AbstractFloat at  C:\Users\...\.julia\packages\GaussianMixtures\3jRIL\src\train.jl:11
davidavdav commented 4 years ago

Sorry that should be 

gmm = GMM(3, X, nIter=10)
giadasp commented 4 years ago

All right! Now it works :)