玩了混合高斯模型,先转几个参考资料,曾经试过自己写代码,结果发现混合高斯模型矩阵运算对我的计算能力要求很高,结果失败了,上网找了代码学习一下大牛们的编程思想,事实证明数学写出来的公式虽然很美,但是现实写代码的时候要考虑各种问题~~~
1.http://www.cnblogs.com/cfantaisie/archive/2011/08/20/2147075.html (主要实现代码)
2.http://freemind.pluskid.org/machine-learning/regularized-gaussian-covariance-estimation/ (对于奇异矩阵的正则化)
3.参考的一本书,Computer Vision Models, Learning, and Inference.(数学推导)
4.斯坦福机器学习课程(数学推导)
利用EM算法:
E-step:
M-step:
matlab实现代码:
function prob = MOF_guassPdf(Data,Mu,Sigma) % % 根据高斯分布函数计算每组数据的概率密度 Probability Density Function (PDF) % 输入 ----------------------------------------------------------------- % o Data: D x N ,N个D维数据 % o Mu: D x 1 ,M个Gauss模型的中心初始值 % o Sigma: D x D ,每个Gauss模型的方差(假设每个方差矩阵都是对角阵, % 即一个数和单位矩阵的乘积) % Outputs ---------------------------------------------------------------- % o prob: N x 1 array representing the probabilities for the % N datapoints. [dim,N] = size(Data); Data = Data' - repmat(Mu',N,1); prob = sum((Data/Sigma).*Data, 2); prob = exp(-0.5*prob) / sqrt((2*pi)^dim * (abs(det(Sigma))+realmin));
EM迭代:
function [Alpha, Mu, Sigma] = MOF_EM(Data, Alpha0, Mu0, Sigma0) %% EM 迭代停止条件 loglik_threshold = 1e-10; %% 初始化参数 [dim, N] = size(Data); M = size(Mu0,2); loglik_old = -realmax; nbStep = 0; %Data是D*N的矩阵 Mu = Mu0; %Mu是D*M的矩阵 Sigma = Sigma0; %sigma是D*D*M的三维矩阵 Alpha = Alpha0; %alpha是1*M大小的向量,意思是列代表第M个高斯模型。 Epsilon = 0.0001; while (nbStep < 1200) nbStep = nbStep+1; %% E-步骤 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% for i=1:M % PDF of each point Pxi(:,i) = MOF_guassPdf(Data, Mu(:,i), Sigma(:,:,i)); end % 计算后验概率 beta(i|x) Pix_tmp = repmat(Alpha,[N 1]).*Pxi; %Pxi应该是N*M的高斯概率矩阵,意思是第N个数据计算第M个高斯函数的概率 Pix = Pix_tmp ./ (repmat(sum(Pix_tmp,2),[1 M])+realmin); Beta = sum(Pix); %% M-步骤 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% for i=1:M % 更新权值 Alpha(i) = Beta(i) / N; % 更新均值 Mu(:,i) = Data*Pix(:,i) / Beta(i); % 更新方差 Data_tmp1 = Data - repmat(Mu(:,i),1,N); Sigma(:,:,i) = (repmat(Pix(:,i)',dim, 1) .* Data_tmp1 * Data_tmp1') / Beta(i); %% Add a tiny variance to avoid numerical instability Sigma(:,:,i) = Sigma(:,:,i) + 1E-5.*diag(ones(dim,1)); end % %% Stopping criterion 1 %%%%%%%%%%%%%%%%%%%% for i=1:M %Compute the new probability p(x|i) Pxi(:,i) = MOF_guassPdf(Data, Mu(:,i), Sigma(i)); end %Compute the log likelihood F = Pxi*Alpha'; F(find(F<realmin)) = realmin; loglik = mean(log(F)); %Stop the process depending on the increase of the log likelihood if abs((loglik/loglik_old)-1) < loglik_threshold break; end loglik_old = loglik; %% Stopping criterion 2 %%%%%%%%%%%%%%%%%%%% %{ v = [sum(abs(Mu - Mu0)), abs(Alpha - Alpha0)]; s = abs(Sigma-Sigma0); v2 = 0; for i=1:M v2 = v2 + det(s(:,:,i)); end if ((sum(v) + v2) < Epsilon) break; end Mu0 = Mu; Sigma0 = Sigma; Alpha0 = Alpha; %} end
测试结果
编写代码随机生成3个高斯分布的数据,参数也随机生成(注意sigma要半正定对称):
这里只画了一个经过EM算法迭代所得参数的高斯函数图..(有点丑,不知道怎么将mesh弄透明,求搜索关键词)。