HMM TOOL

HMM隐马尔科夫模型 MATLAB 工具包对各种数据的处理  

HMM 工具包下载地址：

http://www.cs.ubc.ca/~murphyk/Software/HMM/hmm.html

工具包使用说明：

http://www.cs.ubc.ca/~murphyk/Software/HMM/hmm_usage.html

 

接下来简单叙述一下如何写data

1、data是一维数据、每一组训练样例序列长度一致。

O = 3;
Q = 2; 

prior0 = normalise(rand(Q,1)); 

transmat0 = mk_stochastic(rand(Q,Q));
obsmat0 = mk_stochastic(rand(Q,O));  

%Now we sample nex=20 sequences of length T=10 each from this model, to use as training data.

T=10;
nex=20;
data = dhmm_sample(prior0, transmat0, obsmat0, nex, T);  

%Here data is 20x10. Now we make a random guess as to what the parameters are,

prior1 = normalise(rand(Q,1)); 

transmat1 = mk_stochastic(rand(Q,Q));
obsmat1 = mk_stochastic(rand(Q,O));

%and improve our guess using 5 iterations of EM...

[LL, prior2, transmat2, obsmat2] = dhmm_em(data, prior1, transmat1, obsmat1, 'max_iter', 5);

loglik = dhmm_logprob(data, prior2, transmat2, obsmat2)

%loglik 即用来预测测试数据的相似程度 越大越相似 0为最大

2、data是多维数据、每一组训练样例序列长度一致。

%Let us generate nex=50 vector-valued sequences of length T=50; each vector has size O=2.

O = 2; 

T = 50; 

nex = 50; 

data = randn(O,T,nex);

%Now let use fit a mixture of M=2 Gaussians for each of the Q=2 states using K-means.

M = 2;
Q = 2; 

left_right = 0; 

 prior0 = normalise(rand(Q,1)); 

transmat0 = mk_stochastic(rand(Q,Q)); 

[mu0, Sigma0] = mixgauss_init(Q*M, reshape(data, [O T*nex]), cov_type);

 mu0 = reshape(mu0, [O Q M]); 

Sigma0 = reshape(Sigma0, [O O Q M]);

mixmat0 = mk_stochastic(rand(Q,M));

%Finally, let us improve these parameter estimates using EM.

[LL, prior1, transmat1, mu1, Sigma1, mixmat1] = mhmm_em(data, prior0, transmat0, mu0, Sigma0, mixmat0, 'max_iter', 2);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

说明这里的数组格式是O*T*nex举个例子这个数组是怎么存的

 data0=[x,y,z];data0 是三维数据，供T*nex行，1~T行为nex=1的数据，T+1~2*T行为nex=2的数                %据，以此类推

 data = randn(O,T,nex);

 index=1;

  for k=1:nex

     for j=1:T

         data(:,j,k)=data0(index,:);

         index=index+1;

    end

end 

%按照上述这样将data0写入data即可

%新的数据查看与这个模型的相似程度，即分类

loglik = mhmm_logprob(data, prior, transmat, mu, Sigma, mixmat);

3、data是多维数据、并且每一组训练样例序列长度一致，即HMM如何处理长度不一致数据。

这种情况还是很常见的，例如采集一组连续语音信号，但每次采集得到的长度（帧数）不一致。

假如数据维度为O维，帧数为T（每一组肯能都不一致），NEX为训练数据数目。

步骤1、按照O*T存成NEX行cell类型数据（这里命名为cell_data），例如我的cell_data截图

我的单个数据为8维，供4组训练数据，每一组训练数据取得序列长度不一致。

步骤2、训练代码

    O = 8;%维度

    M = 2;

    Q = 3;

    train_num = 4;

    data =[];

    % initial guess of parameters

   cov_type = 'full';

    % initial guess of parameters

    prior0 = normalise(rand(Q,1));

    transmat0 = mk_stochastic(rand(Q,Q));

    for train_len = 1 : train_num

        data = [data(:, 1 : end), cell_data{train_len}];

    end

    [mu0, Sigma0] = mixgauss_init(Q*M, data, cov_type);

    mu0 = reshape(mu0, [O Q M]);

    Sigma0 = reshape(Sigma0, [O O Q M]);

    mixmat0 = mk_stochastic(rand(Q,M));

    [LL, HMM.prior, HMM.transmat, HMM.mu, HMM.Sigma, HMM.mixmat] = ...