Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Issue #64: added n_init to kmeans #78

Open
wants to merge 8 commits into
base: master
Choose a base branch
from
Open

Issue #64: added n_init to kmeans #78

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
775eefc
Issue #64: added n_init to kmeans
lbollar Sep 30, 2016
f85cce7
cleaned up for PR
lbollar Sep 30, 2016
3ea7c39
removed tabs
lbollar Sep 30, 2016
fd296d4
removed line breaks to conform with style
lbollar Oct 5, 2016
50f0654
fixed another whitespace conflict
lbollar Oct 5, 2016
a9cb7ab
used vim command to hopefully fix trailing whitespace issues
lbollar Oct 5, 2016
b8b381b
extraneous newline
lbollar Oct 5, 2016
5ae2901
Merge branch 'master' into n_init
kmsquire Aug 26, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
53 changes: 34 additions & 19 deletions src/kmeans.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,18 +9,19 @@ type KmeansResult{T<:AbstractFloat} <: ClusteringResult
counts::Vector{Int} # number of samples assigned to each cluster (k)
cweights::Vector{Float64} # cluster weights (k)
totalcost::Float64 # total cost (i.e. objective) (k)
iterations::Int # number of elapsed iterations
iterations::Int # number of elapsed iterations
converged::Bool # whether the procedure converged
end

const _kmeans_default_init = :kmpp
const _kmeans_default_maxiter = 100
const _kmeans_default_tol = 1.0e-6
const _kmeans_default_display = :none
const _kmeans_default_n_init = 10

function kmeans!{T<:AbstractFloat}(X::Matrix{T}, centers::Matrix{T};
weights=nothing,
maxiter::Integer=_kmeans_default_maxiter,
maxiter::Integer=_kmeans_default_maxiter,
tol::Real=_kmeans_default_tol,
display::Symbol=_kmeans_default_display)

Expand All @@ -34,27 +35,41 @@ function kmeans!{T<:AbstractFloat}(X::Matrix{T}, centers::Matrix{T};
counts = Array(Int, k)
cweights = Array(Float64, k)

_kmeans!(X, conv_weights(T, n, weights), centers,
assignments, costs, counts, cweights,
_kmeans!(X, conv_weights(T, n, weights), centers,
assignments, costs, counts, cweights,
round(Int, maxiter), tol, display_level(display))
end

function kmeans(X::Matrix, k::Int;
function kmeans(X::Matrix, k::Int;
weights=nothing,
init=_kmeans_default_init,
maxiter::Integer=_kmeans_default_maxiter,
maxiter::Integer=_kmeans_default_maxiter,
n_init::Integer=_kmeans_default_n_init,
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I understand that n_init comes from Python's sklearn (#64), but it doesn't sound like a best choice for me.
Maybe something like n_tries to reflect that the parameter defines how many times the algorithm, rather than some initialization procedure, is run?

Copy link
Contributor

@wildart wildart Sep 28, 2018 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

or ntries? And wouldn't be an overkill to run 10 times? I recommend default value 1, because usually a quick partitioning is required and not necessarily best one. And, if one needs to find a best clustering, this parameter can be set to larger value explicitly.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

10 is what sklearn does at it sounds reasonable to me.
It isn't unusual to run 1000s of times, (that was done as the baseline for the affinity propagation paper)
If some need a quick partition they can ask for it.

The default shouldn't be so sensitive to random factors.

I think 10 strikes the right balance.
Though I could see argument for 3 or 30

tol::Real=_kmeans_default_tol,
display::Symbol=_kmeans_default_display)

m, n = size(X)
(2 <= k < n) || error("k must have 2 <= k < n.")
iseeds = initseeds(init, X, k)
centers = copyseeds(X, iseeds)
kmeans!(X, centers;
weights=weights,
maxiter=maxiter,
tol=tol,
display=display)
n_init > 0 || throw(ArgumentError("n_init must be greater than 0"))

lowestcost::Float64 = Inf
local bestresult::KmeansResult

for i = 1:n_init
iseeds = initseeds(init, X, k)
centers = copyseeds(X, iseeds)
result = kmeans!(X, centers;
weights=weights,
maxiter=maxiter,
tol=tol,
display=display)

if result.totalcost < lowestcost
lowestcost = result.totalcost
bestresult = result
end
end
return bestresult
end

#### Core implementation
Expand All @@ -68,8 +83,8 @@ function _kmeans!{T<:AbstractFloat}(
costs::Vector{T}, # out: costs of the resultant assignments (n)
counts::Vector{Int}, # out: the number of samples assigned to each cluster (k)
cweights::Vector{Float64}, # out: the weights of each cluster
maxiter::Int, # in: maximum number of iterations
tol::Real, # in: tolerance of change at convergence
maxiter::Int, # in: maximum number of iterations
tol::Real, # in: tolerance of change at convergence
displevel::Int) # in: the level of display

# initialize
Expand Down Expand Up @@ -153,7 +168,7 @@ function _kmeans!{T<:AbstractFloat}(
end
end

return KmeansResult(centers, assignments, costs, counts, cweights,
return KmeansResult(centers, assignments, costs, counts, cweights,
@compat(Float64(objv)), t, converged)
end

Expand Down Expand Up @@ -245,7 +260,7 @@ function update_centers!{T<:AbstractFloat}(
n::Int = size(x, 2)
k::Int = size(centers, 2)

# initialize center weights
# initialize center weights
for i = 1 : k
if to_update[i]
cweights[i] = 0.
Expand Down Expand Up @@ -299,7 +314,7 @@ function update_centers!{T<:AbstractFloat}(
n::Int = size(x, 2)
k::Int = size(centers, 2)

# initialize center weights
# initialize center weights
for i = 1 : k
if to_update[i]
cweights[i] = 0.
Expand All @@ -314,7 +329,7 @@ function update_centers!{T<:AbstractFloat}(
if wj > 0
@inbounds cj = assignments[j]
1 <= cj <= k || error("assignment out of boundary.")

if to_update[cj]
rj = view(centers, :, cj)
xj = view(x, :, j)
Expand Down
6 changes: 3 additions & 3 deletions test/kmeans.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ k = 10
x = rand(m, n)

# non-weighted
r = kmeans(x, k; maxiter=50)
r = kmeans(x, k; maxiter=50, n_init=2)
@test isa(r, KmeansResult{Float64})
@test size(r.centers) == (m, k)
@test length(r.assignments) == n
Expand All @@ -24,7 +24,7 @@ r = kmeans(x, k; maxiter=50)
@test_approx_eq sum(r.costs) r.totalcost

# non-weighted (float32)
r = kmeans(@compat(map(Float32, x)), k; maxiter=50)
r = kmeans(@compat(map(Float32, x)), k; maxiter=50, n_init=2)
@test isa(r, KmeansResult{Float32})
@test size(r.centers) == (m, k)
@test length(r.assignments) == n
Expand All @@ -37,7 +37,7 @@ r = kmeans(@compat(map(Float32, x)), k; maxiter=50)

# weighted
w = rand(n)
r = kmeans(x, k; maxiter=50, weights=w)
r = kmeans(x, k; maxiter=50, weights=w, n_init=2)
@test isa(r, KmeansResult{Float64})
@test size(r.centers) == (m, k)
@test length(r.assignments) == n
Expand Down