Trainer that uses the Baum-Welch algorithm to estimate parameters.
These should be used when a training sequence for an HMM has unknown
paths for the actual states, and you need to make an estimation of the
model parameters from the observed emissions.
This uses the Baum-Welch algorithm, first described in
Baum, L.E. 1972. Inequalities. 3:1-8
This is based on the description in Biological Sequence Analysis by
Durbin et al. in section 3.3
This algorithm is guaranteed to converge to a local maximum, but not
necessarily to the global maxima, so use with care!
|
Methods
|
|
__init__
train
update_emissions
update_transitions
|
|
|
__init__
|
__init__ ( self, markov_model )
Initialize the trainer.
Arguments:
|
|
|
train
|
train (
self,
training_seqs,
stopping_criteria,
dp_method=ScaledDPAlgorithms,
)
Estimate the parameters using training sequences.
The algorithm for this is taken from Durbin et al. p64, so this
is a good place to go for a reference on what is going on.
Arguments:
training_seqs -- A list of TrainingSequence objects to be used
for estimating the parameters. stopping_criteria -- A function, that when passed the change
in log likelihood and threshold, will indicate if we should stop
the estimation iterations. dp_method -- A class instance specifying the dynamic programming
implementation we should use to calculate the forward and
backward variables. By default, we use the scaling method.
|
|
|
update_emissions
|
update_emissions (
self,
emission_counts,
training_seq,
forward_vars,
backward_vars,
training_seq_prob,
)
Add the contribution of a new training sequence to the emissions
Arguments:
emission_counts -- A dictionary of the current counts for the
emissions training_seq -- The training sequence we are working with forward_vars -- Probabilities calculated using the forward
algorithm. backward_vars -- Probabilities calculated using the backwards
algorithm. training_seq_prob - The probability of the current sequence.
This calculates E_{k}(b) (the estimated emission probability for
emission letter b from state k) using formula 3.21 in Durbin et al.
|
|
|
update_transitions
|
update_transitions (
self,
transition_counts,
training_seq,
forward_vars,
backward_vars,
training_seq_prob,
)
Add the contribution of a new training sequence to the transitions.
Arguments:
transition_counts -- A dictionary of the current counts for the
transitions training_seq -- The training sequence we are working with forward_vars -- Probabilities calculated using the forward
algorithm. backward_vars -- Probabilities calculated using the backwards
algorithm. training_seq_prob - The probability of the current sequence.
This calculates A_{kl} (the estimated transition counts from state
k to state l) using formula 3.20 in Durbin et al.
|
|