Daniel@0: function [LL, prior, transmat, obsmat, nrIterations] = ...
Daniel@0:    dhmm_em(data, prior, transmat, obsmat, varargin)
Daniel@0: % LEARN_DHMM Find the ML/MAP parameters of an HMM with discrete outputs using EM.
Daniel@0: % [ll_trace, prior, transmat, obsmat, iterNr] = learn_dhmm(data, prior0, transmat0, obsmat0, ...)
Daniel@0: %
Daniel@0: % Notation: Q(t) = hidden state, Y(t) = observation
Daniel@0: %
Daniel@0: % INPUTS:
Daniel@0: % data{ex} or data(ex,:) if all sequences have the same length
Daniel@0: % prior(i)
Daniel@0: % transmat(i,j)
Daniel@0: % obsmat(i,o)
Daniel@0: %
Daniel@0: % Optional parameters may be passed as 'param_name', param_value pairs.
Daniel@0: % Parameter names are shown below; default values in [] - if none, argument is mandatory.
Daniel@0: %
Daniel@0: % 'max_iter' - max number of EM iterations [10]
Daniel@0: % 'thresh' - convergence threshold [1e-4]
Daniel@0: % 'verbose' - if 1, print out loglik at every iteration [1]
Daniel@0: % 'obs_prior_weight' - weight to apply to uniform dirichlet prior on observation matrix [0]
Daniel@0: %
Daniel@0: % To clamp some of the parameters, so learning does not change them:
Daniel@0: % 'adj_prior' - if 0, do not change prior [1]
Daniel@0: % 'adj_trans' - if 0, do not change transmat [1]
Daniel@0: % 'adj_obs' - if 0, do not change obsmat [1]
Daniel@0: %
Daniel@0: % Modified by Herbert Jaeger so xi are not computed individually
Daniel@0: % but only their sum (over time) as xi_summed; this is the only way how they are used
Daniel@0: % and it saves a lot of memory.
Daniel@0: 
Daniel@0: [max_iter, thresh, verbose, obs_prior_weight, adj_prior, adj_trans, adj_obs] = ...
Daniel@0:    process_options(varargin, 'max_iter', 10, 'thresh', 1e-4, 'verbose', 1, ...
Daniel@0:                    'obs_prior_weight', 0, 'adj_prior', 1, 'adj_trans', 1, 'adj_obs', 1);
Daniel@0: 
Daniel@0: previous_loglik = -inf;
Daniel@0: loglik = 0;
Daniel@0: converged = 0;
Daniel@0: num_iter = 1;
Daniel@0: LL = [];
Daniel@0: 
Daniel@0: if ~iscell(data)
Daniel@0:  data = num2cell(data, 2); % each row gets its own cell
Daniel@0: end
Daniel@0: 
Daniel@0: while (num_iter <= max_iter) & ~converged
Daniel@0:  % E step
Daniel@0:  [loglik, exp_num_trans, exp_num_visits1, exp_num_emit] = ...
Daniel@0:      compute_ess_dhmm(prior, transmat, obsmat, data, obs_prior_weight);
Daniel@0: 
Daniel@0:  % M step
Daniel@0:  if adj_prior
Daniel@0:    prior = normalise(exp_num_visits1);
Daniel@0:  end
Daniel@0:  if adj_trans & ~isempty(exp_num_trans)
Daniel@0:    transmat = mk_stochastic(exp_num_trans);
Daniel@0:  end
Daniel@0:  if adj_obs
Daniel@0:    obsmat = mk_stochastic(exp_num_emit);
Daniel@0:  end
Daniel@0: 
Daniel@0:  if verbose, fprintf(1, 'iteration %d, loglik = %f\n', num_iter, loglik); end
Daniel@0:  num_iter =  num_iter + 1;
Daniel@0:  converged = em_converged(loglik, previous_loglik, thresh);
Daniel@0:  previous_loglik = loglik;
Daniel@0:  LL = [LL loglik];
Daniel@0: end
Daniel@0: nrIterations = num_iter - 1;
Daniel@0: 
Daniel@0: %%%%%%%%%%%%%%%%%%%%%%%
Daniel@0: 
Daniel@0: function [loglik, exp_num_trans, exp_num_visits1, exp_num_emit, exp_num_visitsT] = ...
Daniel@0:    compute_ess_dhmm(startprob, transmat, obsmat, data, dirichlet)
Daniel@0: % COMPUTE_ESS_DHMM Compute the Expected Sufficient Statistics for an HMM with discrete outputs
Daniel@0: % function [loglik, exp_num_trans, exp_num_visits1, exp_num_emit, exp_num_visitsT] = ...
Daniel@0: %    compute_ess_dhmm(startprob, transmat, obsmat, data, dirichlet)
Daniel@0: %
Daniel@0: % INPUTS:
Daniel@0: % startprob(i)
Daniel@0: % transmat(i,j)
Daniel@0: % obsmat(i,o)
Daniel@0: % data{seq}(t)
Daniel@0: % dirichlet - weighting term for uniform dirichlet prior on expected emissions
Daniel@0: %
Daniel@0: % OUTPUTS:
Daniel@0: % exp_num_trans(i,j) = sum_l sum_{t=2}^T Pr(X(t-1) = i, X(t) = j| Obs(l))
Daniel@0: % exp_num_visits1(i) = sum_l Pr(X(1)=i | Obs(l))
Daniel@0: % exp_num_visitsT(i) = sum_l Pr(X(T)=i | Obs(l))
Daniel@0: % exp_num_emit(i,o) = sum_l sum_{t=1}^T Pr(X(t) = i, O(t)=o| Obs(l))
Daniel@0: % where Obs(l) = O_1 .. O_T for sequence l.
Daniel@0: 
Daniel@0: numex = length(data);
Daniel@0: [S O] = size(obsmat);
Daniel@0: exp_num_trans = zeros(S,S);
Daniel@0: exp_num_visits1 = zeros(S,1);
Daniel@0: exp_num_visitsT = zeros(S,1);
Daniel@0: exp_num_emit = dirichlet*ones(S,O);
Daniel@0: loglik = 0;
Daniel@0: 
Daniel@0: for ex=1:numex
Daniel@0:  obs = data{ex};
Daniel@0:  T = length(obs);
Daniel@0:  %obslik = eval_pdf_cond_multinomial(obs, obsmat);
Daniel@0:  obslik = multinomial_prob(obs, obsmat);
Daniel@0:  [alpha, beta, gamma, current_ll, xi_summed] = fwdback(startprob, transmat, obslik);
Daniel@0: 
Daniel@0:  loglik = loglik +  current_ll;
Daniel@0:  exp_num_trans = exp_num_trans + xi_summed;
Daniel@0:  exp_num_visits1 = exp_num_visits1 + gamma(:,1);
Daniel@0:  exp_num_visitsT = exp_num_visitsT + gamma(:,T);
Daniel@0:  % loop over whichever is shorter
Daniel@0:  if T < O
Daniel@0:    for t=1:T
Daniel@0:      o = obs(t);
Daniel@0:      exp_num_emit(:,o) = exp_num_emit(:,o) + gamma(:,t);
Daniel@0:    end
Daniel@0:  else
Daniel@0:    for o=1:O
Daniel@0:      ndx = find(obs==o);
Daniel@0:      if ~isempty(ndx)
Daniel@0:        exp_num_emit(:,o) = exp_num_emit(:,o) + sum(gamma(:, ndx), 2);
Daniel@0:      end
Daniel@0:    end
Daniel@0:  end
Daniel@0: end