使用EM算法,我想在给定数据集上训练具有四个分量的高斯混合模型。该集是三维的,包含300个样本。
问题在于,在大约6轮EM算法之后,根据matlab(rank(sigma) = 2
而不是3),协方差矩阵sigma变得接近于奇异。这反过来会导致不希望的结果,例如评估高斯分布的复杂值gm(k,i)
。
此外,我使用高斯的对数来解释下溢问题 - 参见E步骤。我不确定这是否正确,如果我必须在其他地方采取责任p(w_k | x ^(i),theta)的exp?
你能告诉我到目前为止我的EM算法实现是否正确吗? 如何解决接近奇异协方差西格玛的问题?
以下是我对EM算法的实现:
首先,我使用kmeans 初始化组件的均值和协方差:
load('data1.mat');
X = Data'; % 300x3 data set
D = size(X,2); % dimension
N = size(X,1); % number of samples
K = 4; % number of Gaussian Mixture components
% Initialization
p = [0.2, 0.3, 0.2, 0.3]; % arbitrary pi
[idx,mu] = kmeans(X,K); % initial means of the components
% compute the covariance of the components
sigma = zeros(D,D,K);
for k = 1:K
sigma(:,:,k) = cov(X(idx==k,:));
end
对于 E-step ,我使用以下公式计算责任。
w_k是k高斯分量。
x ^(i)是单个数据点(样本)
theta代表高斯混合模型的参数:mu,Sigma,pi。
以下是相应的代码:
% variables for convergence
converged = 0;
prevLoglikelihood = Inf;
prevMu = mu;
prevSigma = sigma;
prevPi = p;
round = 0;
while (converged ~= 1)
round = round +1
gm = zeros(K,N); % gaussian component in the nominator
sumGM = zeros(N,1); % denominator of responsibilities
% E-step: Evaluate the responsibilities using the current parameters
% compute the nominator and denominator of the responsibilities
for k = 1:K
for i = 1:N
Xmu = X-mu;
% I am using log to prevent underflow of the gaussian distribution (exp("small value"))
logPdf = log(1/sqrt(det(sigma(:,:,k))*(2*pi)^D)) + (-0.5*Xmu*(sigma(:,:,k)\Xmu'));
gm(k,i) = log(p(k)) * logPdf;
sumGM(i) = sumGM(i) + gm(k,i);
end
end
% calculate responsibilities
res = zeros(K,N); % responsibilities
Nk = zeros(4,1);
for k = 1:K
for i = 1:N
% I tried to use the exp(gm(k,i)/sumGM(i)) to compute res but this leads to sum(pi) > 1.
res(k,i) = gm(k,i)/sumGM(i);
end
Nk(k) = sum(res(k,:));
end
Nk(k)
使用M步骤中给出的公式计算,并在M步骤中用于计算新概率p(k)
。
M-步骤
% M-step: Re-estimate the parameters using the current responsibilities
for k = 1:K
for i = 1:N
mu(k,:) = mu(k,:) + res(k,i).*X(k,:);
sigma(:,:,k) = sigma(:,:,k) + res(k,i).*(X(k,:)-mu(k,:))*(X(k,:)-mu(k,:))';
end
mu(k,:) = mu(k,:)./Nk(k);
sigma(:,:,k) = sigma(:,:,k)./Nk(k);
p(k) = Nk(k)/N;
end
现在,为了检查收敛性,使用以下公式计算对数似然:
% Evaluate the log-likelihood and check for convergence of either
% the parameters or the log-likelihood. If not converged, go to E-step.
loglikelihood = 0;
for i = 1:N
loglikelihood = loglikelihood + log(sum(gm(:,i)));
end
% Check for convergence of parameters
errorLoglikelihood = abs(loglikelihood-prevLoglikelihood);
if (errorLoglikelihood <= eps)
converged = 1;
end
errorMu = abs(mu(:)-prevMu(:));
errorSigma = abs(sigma(:)-prevSigma(:));
errorPi = abs(p(:)-prevPi(:));
if (all(errorMu <= eps) && all(errorSigma <= eps) && all(errorPi <= eps))
converged = 1;
end
prevLoglikelihood = loglikelihood;
prevMu = mu;
prevSigma = sigma;
prevPi = p;
end % while
我的Matlab实现高斯混合模型的EM算法有什么问题吗?
以前的麻烦:
问题是我无法使用对数似然检查收敛,因为它是-Inf
。这是在舍入零值的同时评估责任公式中的高斯值(参见E步骤)。
你能告诉我到目前为止我的EM算法实现是否正确吗? 以及如何使用舍入的零值来解决问题?
以下是我对EM算法的实现:
首先,我使用kmeans 初始化组件的均值和协方差:
load('data1.mat');
X = Data'; % 300x3 data set
D = size(X,2); % dimension
N = size(X,1); % number of samples
K = 4; % number of Gaussian Mixture components
% Initialization
p = [0.2, 0.3, 0.2, 0.3]; % arbitrary pi
[idx,mu] = kmeans(X,K); % initial means of the components
% compute the covariance of the components
sigma = zeros(D,D,K);
for k = 1:K
sigma(:,:,k) = cov(X(idx==k,:));
end
以下是相应的代码:
% variables for convergence
converged = 0;
prevLoglikelihood = Inf;
prevMu = mu;
prevSigma = sigma;
prevPi = p;
round = 0;
while (converged ~= 1)
round = round +1
gm = zeros(K,N); % gaussian component in the nominator -
% some values evaluate to zero
sumGM = zeros(N,1); % denominator of responsibilities
% E-step: Evaluate the responsibilities using the current parameters
% compute the nominator and denominator of the responsibilities
for k = 1:K
for i = 1:N
% HERE values evalute to zero e.g. exp(-746.6228) = -Inf
gm(k,i) = p(k)/sqrt(det(sigma(:,:,k))*(2*pi)^D)*exp(-0.5*(X(i,:)-mu(k,:))*inv(sigma(:,:,k))*(X(i,:)-mu(k,:))');
sumGM(i) = sumGM(i) + gm(k,i);
end
end
% calculate responsibilities
res = zeros(K,N); % responsibilities
Nk = zeros(4,1);
for k = 1:K
for i = 1:N
res(k,i) = gm(k,i)/sumGM(i);
end
Nk(k) = sum(res(k,:));
end
Nk(k)
使用M步骤中给出的公式计算。
M-步骤
% M-step: Re-estimate the parameters using the current responsibilities
mu = zeros(K,3);
for k = 1:K
for i = 1:N
mu(k,:) = mu(k,:) + res(k,i).*X(k,:);
sigma(:,:,k) = sigma(:,:,k) + res(k,i).*(X(k,:)-mu(k,:))*(X(k,:)-mu(k,:))';
end
mu(k,:) = mu(k,:)./Nk(k);
sigma(:,:,k) = sigma(:,:,k)./Nk(k);
p(k) = Nk(k)/N;
end
% Evaluate the log-likelihood and check for convergence of either
% the parameters or the log-likelihood. If not converged, go to E-step.
loglikelihood = 0;
for i = 1:N
loglikelihood = loglikelihood + log(sum(gm(:,i)));
end
% Check for convergence of parameters
errorLoglikelihood = abs(loglikelihood-prevLoglikelihood);
if (errorLoglikelihood <= eps)
converged = 1;
end
errorMu = abs(mu(:)-prevMu(:));
errorSigma = abs(sigma(:)-prevSigma(:));
errorPi = abs(p(:)-prevPi(:));
if (all(errorMu <= eps) && all(errorSigma <= eps) && all(errorPi <= eps))
converged = 1;
end
prevLoglikelihood = loglikelihood;
prevMu = mu;
prevSigma = sigma;
prevPi = p;
end % while
第一轮结束后loglikelihood
约为700。
在第二轮中,它是-Inf
,因为E步骤中的一些gm(k,i)
值为零。因此,对数显然是负无穷大。
零值也会导致sumGM
等于零,从而导致mu
和sigma
矩阵内的所有NaN条目。
我该如何解决这个问题? 你能告诉我我的实施是否有问题吗? 可以通过某种方式提高Matlab的精度来解决吗?
修改
我在gm(k,i)中添加了exp()项的缩放。 不幸的是,这并没有多大帮助。经过多次回合后,我仍然遇到了下溢问题。
scale = zeros(N,D);
for i = 1:N
max = 0;
for k = 1:K
Xmu = X(i,:)-mu(k,:);
if (norm(scale(i,:) - Xmu) > max)
max = norm(scale(i,:) - Xmu);
scale(i,:) = Xmu;
end
end
end
for k = 1:K
for i = 1:N
Xmu = X(i,:)-mu(k,:);
% scale gm to prevent underflow
Xmu = Xmu - scale(i,:);
gm(k,i) = p(k)/sqrt(det(sigma(:,:,k))*(2*pi)^D)*exp(-0.5*Xmu*inv(sigma(:,:,k))*Xmu');
sumGM(i) = sumGM(i) + gm(k,i);
end
end
此外,我注意到kmeans初始化的方法与下一轮相比完全不同,其中平均值在M步骤中计算。
k均值:
mu = 13.500000000000000 0.026602138870044 0.062415945993735
88.500000000000000 -0.009869960132085 -0.075177888210981
39.000000000000000 -0.042569305020309 0.043402772876513
64.000000000000000 -0.024519281362918 -0.012586980924762
在M步之后:
round = 2
mu = 1.000000000000000 0.077230046948357 0.024498886414254
2.000000000000000 0.074260118474053 0.026484346404660
3.000000000000002 0.070944016105476 0.029043085983168
4.000000000000000 0.067613431480832 0.031641849205021
在下一轮mu
根本没有变化。它与第2轮保持不变。
我猜这是因为gm(k,i)的下溢引起的? 要么我的缩放实现不正确,要么算法的整个实现在某处错误:(
编辑2
经过四轮比赛后,我获得了NaN
个值,并更详细地研究了gm。仅查看一个样本(并且没有0.5因子),gm
在所有组件中变为零。放入matlab gm(:,1) = [0 0 0 0]
。这反过来导致sumGM等于零 - &gt; NaN,因为我除以零。我在
round = 1
mu = 62.0000 -0.0298 -0.0078
37.0000 -0.0396 0.0481
87.5000 -0.0083 -0.0728
12.5000 0.0303 0.0614
gm(:,1) = [11.7488, 0.0000, 0.0000, 0.0000]
round = 2
mu = 1.0000 0.0772 0.0245
2.0000 0.0743 0.0265
3.0000 0.0709 0.0290
4.0000 0.0676 0.0316
gm(:,1) = [0.0000, 0.0000, 0.0000, 0.3128]
round = 3
mu = 1.0000 0.0772 0.0245
2.0000 0.0743 0.0265
3.0000 0.0709 0.0290
4.0000 0.0676 0.0316
gm(:,1) = [0, 0, 0.0000, 0.2867]
round = 4
mu = 1.0000 0.0772 0.0245
NaN NaN NaN
3.0000 0.0709 0.0290
4.0000 0.0676 0.0316
gm(:,1) = 1.0e-105 * [0, NaN, 0, 0.5375]
首先,手段似乎没有变化,与kmeans的初始化相比完全不同。
根据gm(:,1)
的输出,每个样本(不仅仅是像这里的第一个样本)仅对应于一个高斯分量。不应该将样本部分分发&#34;在每个高斯组件中?
EDIT3:
所以我猜mu不改变的问题是M步骤的第一行:mu = zeros(K,3);
。
为了解决下溢问题,我目前正在尝试使用高斯的日志:
function logPdf = logmvnpdf(X, mu, sigma, D)
Xmu = X-mu;
logPdf = log(1/sqrt(det(sigma)*(2*pi)^D)) + (-0.5*Xmu*inv(sigma)*Xmu');
end
新问题是协方差矩阵sigma。 Matlab声称: 警告:矩阵接近单一或严重缩放。结果可能不准确。
经过6轮后,我得到了gm(高斯分布)的虚数值。
更新的E-Step现在看起来像这样:
gm = zeros(K,N); % gaussian component in the nominator
sumGM = zeros(N,1); % denominator of responsibilities
for k = 1:K
for i = 1:N
%gm(k,i) = p(k)/sqrt(det(sigma(:,:,k))*(2*pi)^D)*exp(-0.5*Xmu*inv(sigma(:,:,k))*Xmu');
%gm(k,i) = p(k)*mvnpdf(X(i,:),mu(k,:),sigma(:,:,k));
gm(k,i) = log(p(k)) + logmvnpdf(X(i,:), mu(k,:), sigma(:,:,k), D);
sumGM(i) = sumGM(i) + gm(k,i);
end
end
答案 0 :(得分:3)
看起来你应该能够使用比例因子标度(i)将gm(k,i)带入可表示的范围,因为如果你按比例(i)乘以gm(k,i),这将结束也可以将sumGM(i)相乘,并在计算出res(k,i)= gm(k,i)/ sumGM(i)时取消。
理论上我会使scale(i)= 1 / max_k(exp(-0.5 *(X(i,:) - mu(k,:)))并且实际上在不进行求幂的情况下进行计算,所以你最终处理它的日志,max_k(-0.5 *(X(i,:)) - mu(k,:)) - 这给你一个可以添加到每个-0.5 *(X(i,:))的通用术语 - 在使用exp()之前mu(k,:)并且将至少保持在可表示范围内的最大值 - 在此更正之后仍然下溢到零的任何事情你都不关心,因为它与其他贡献相比非常小