Massive changes to the training algorithms, including weighted states, edge inertia, pseudocounts on edges, and uniform pseudocounts added. Viterbi has been cleaned up significantly by having it be a special case of the more abstract training on labelled sequences method.
Other misc changes include adding the Model.log_probability() method, changing forward-backward emissions to optionally be tied, and transitions to give the expected number of transitions, and not the log probability. A MultivariateDistribution distribution has been added, for arbitrary numbers of independent features.
Massive changes to the training algorithms, including weighted states, edge inertia, pseudocounts on edges, and uniform pseudocounts added. Viterbi has been cleaned up significantly by having it be a special case of the more abstract training on labelled sequences method.
Other misc changes include adding the Model.log_probability() method, changing forward-backward emissions to optionally be tied, and transitions to give the expected number of transitions, and not the log probability. A MultivariateDistribution distribution has been added, for arbitrary numbers of independent features.