classdef Kmeans < handle & Learner

%TODO : lite seems to be better? change to formal setting in the future?

% kihyuk's library has a pretty nice function

properties

weights; %feadim*numunits

feadim;

numunits;

type; %activation type

threshold = 0.1; %for type 'soft' (Adam Coaste used 0.1, 0.25, 0.5, 1.0)

%need to clean up these in the future

savepath;

E;

end

methods

function self = Kmeans(N, type)

addpath(genpath(fullfile(Config.lib_dir_path,'minFunc')));

self.numunits = N;

self.type = type;

end

function train(self, X)

%deal with empty center better

%re-implement this later

% if ~isempty(savedir_s)

% savedir = sprintf('/mnt/neocortex/scratch/suii/basis/%s',savedir_s);

% if exist(savedir,'dir')

% disp('warning : directory already exist');

% else

% mkdir(savedir);

% end

% end

self.feadim = size(X,1);

[~, center] = Kmeans.litekmeans(X, self.numunits, true, self.max_iter);

self.weights = center;

% if ~isempty(savedir_s)

% learner_id = sprintf('Kmeans_%s_N%g',self.type,self.numunits);

% fname_save = sprintf('%s/%s', savedir , learner_id);

% fname_mat = sprintf('%s.mat', fname_save);

% fname_png_1 = sprintf('%s_1.png', fname_save);

% self.savepath = [savedir_s , '_' , learner_id];

% DeepBeliefNetwork.save_progress(self, fname_mat, fname_png_1);

% end

end

% function train_old(self, savedir_s, Data, debug)

% X = Data.Xtrain;

% self.numchannels = Data.numchannels;

% self.prev_learner = Data.learner;

% clear Data;

% savedir = sprintf('/mnt/neocortex/scratch/suii/basis/%s',savedir_s);

% if exist(savedir,'dir')

% disp('warning : directory already exist');

% else

% mkdir(savedir);

% end

% X = X';

% k = self.numunits;

% x2 = sum(X.^2,2);

% centroids = randn(k,size(X,2))*0.1; %X(randsample(size(X,1), k), :);

% BATCH_SIZE=1000;

% iterations = 50;

% loss_history = zeros(iterations,1);

% if exist('debug','var') && ~isempty(debug) && debug

% iterations = 1;

% end

% for itr = 1:iterations

% fprintf('K-means iteration %d / %d\n', itr, iterations);

% c2 = 0.5*sum(centroids.^2,2);

% summation = zeros(k, size(X,2));

% counts = zeros(k, 1);

% loss =0;

% for i=1:BATCH_SIZE:size(X,1)

% lastIndex=min(i+BATCH_SIZE-1, size(X,1));

% m = lastIndex - i + 1;

% [val,labels] = max(bsxfun(@minus,centroids*X(i:lastIndex,:)',c2));

% loss = loss + sum(0.5*x2(i:lastIndex) - val');

% S = sparse(1:m,labels,1,m,k,m); % labels as indicator matrix

% summation = summation + S'*X(i:lastIndex,:);

% counts = counts + sum(S,1)';

% end

% loss_history(itr) = loss;

% centroids = bsxfun(@rdivide, summation, counts);

% just zap empty centroids so they don't introduce NaNs everywhere.

% badIndex = find(counts == 0);

% centroids(badIndex, :) = 0;

% end

% self.weights = centroids';

% learner_id = sprintf('Kmeans_%s_N%g',self.type,self.numunits);

% fname_save = sprintf('%s/%s', savedir , learner_id);

% fname_mat = sprintf('%s.mat', fname_save);

% fname_png_1 = sprintf('%s_1.png', fname_save);

% fname_png_2 = sprintf('%s_2.png', fname_save);

% self.savepath = [savedir_s , '_' , learner_id];

% DeepBeliefNetwork.save_progress(self, fname_mat, fname_png_1,fname_png_2, {loss_history}, {'loss_history'});

% end

function [acti] = fprop(self, patches)

numsamples = size(patches,2);

patches = patches';

centroids = self.weights';

xx = sum(patches.^2, 2);

cc = sum(centroids.^2, 2)';

xc = patches * centroids'; %numsamples * numunits

z = real(sqrt( bsxfun(@plus, cc, bsxfun(@minus, xx, 2*xc)) )); % distances, real for numerical error

switch self.type

case 'tri'

mu = mean(z, 2); % average distance to centroids for each patch

acti = max(bsxfun(@minus, mu, z), 0)';

case 'soft'

acti = max(0,xc-self.threshold)';

case 'hard'

[v, inds] = min(z,[],2);

acti = zeros(self.numunits, numsamples);

idx = full(sparse([1:numsamples]', inds, ones(numsamples,1 ) ))';

acti(1:size(idx,1),:) = idx;

end

end

function [weights] = show_basis(self,dum)

weights = self.weights;

end

end

methods(Static)

function [label,center] = litekmeans(X, k, opt_verbose, MAX_ITERS)

if ~exist('opt_verbose', 'var')

opt_verbose = false;

end

if ~exist('MAX_ITERS', 'var')

MAX_ITERS = 50;

end

n = size(X,2);

last = 0;

label = ceil(k*rand(1,n)); % random initialization

itr=0;

% MAX_ITERS=50;

while any(label ~= last)

itr = itr+1;

if opt_verbose

fprintf( '%d(%d)..', itr, sum(label ~= last));

end

E = sparse(1:n,label,1,n,k,n); % transform label into indicator matrix

center = X*(E*spdiags(1./sum(E,1)',0,k,k)); % compute center of each cluster

last = label;

[val,label] = max(bsxfun(@minus,center'*X,0.5*sum(center.^2,1)')); % assign samples to the nearest centers

if (itr >= MAX_ITERS) break; end;

end

% center=center';

if opt_verbose

fprintf('\n');

end

end

end