Class BaumWelchTrainer

x.__init__(...) initializes x; see help(type(x)) for signature

Overrides: object.__init__

(inherited documentation)

Verifies and processes the training option values. Returns the processed dict.

Stores a line in the history of the model or wherever else it is appropriate to keep a record of training steps.

Default implementation does nothing, but subclasses may want to store this information.

This should be overridden by subclasses, but not by defining a static method on the class, since the function must be picklable. For this, it needs to be a top-level function. Then you can set the sequence_updates attribute to point to it (using staticmethod), as we have done in the default implementation.

Parameters:

• update_initial (bool) - usually you want to update all distributions, including the initial state distribution. If update_initial=False, the initial state distribution updates won't be made for this sequence. We want this when the sequence is actually a non-initial fragment of a longer sequence

Creates a mutable version of the given model. This mutable version will be the model that receives updates during training, as defined by update_model.

Creates empty arrays to hold the accumulated probabilities during training. The sizes will depend on self.model.

Returns a tuple of the dicts that map labels, emissions, etc to the indices of arrays to which they correspond. These will need to be different for non-standard models.

Callback for the sequence_updates processes that takes the updates from a single sequence and adds them onto the global update accumulators.

The accumulators are stored as self.global_arrays.

Replaces the distributions of the saved model with the probabilities taken from the arrays of updates. self.model is expected to be made up of mutable distributions when this is called.

Saves the model in self.model to disk. This may be called at the end of each iteration and will be called at the end of the whole training process.

By default, does nothing. You don't have to put something in here, but you'll need to override this if you want the model to be saved during training before it gets return at the end.

Performs unsupervised training using Baum-Welch EM.

This is performed as a retraining step on a model that has already been initialized.

This is based on the training procedure in NLTK for HMMs: nltk.tag.hmm.HiddenMarkovModelTrainer.train_unsupervised.

Parameters:

• emissions (list of lists of emissions) - training data. Each element is a list of emissions representing a sequence in the training data. Each emission is an emission like those used for emission_log_probability on the model
• logger (logging.Logger) - a logger to send progress logging to

OPTIONS

Value:

[ModuleOption('max_iterations', filter= int, help_text= "Number of tra ining iterations to give up after " "if we don't reach convergence bef ore.", usage= "max_iterations=N, where N is an integer", default= 100) , ModuleOption('convergence_logprob', filter= float, help_text= "Diffe rence in overall log probability of the " "training data made by one i teration after which we " "consider the training to have converged.", usage= "convergence_logprob=X, where X is a small floating " "point nu mber (e.g. 1e-3)", default= 1e-3), ModuleOption('split', filter= int, ...