API Reference

class pycrfsuite.ItemSequence

A wrapper for crfsuite ItemSequence - a class for storing features for all items in a single sequence.

Using this class is an alternative to passing data to Trainer and Tagger directly. By using this class it is possible to save some time if the same input sequence is passed to trainers/taggers more than once - features won’t be processed multiple times. It also allows to get “processed” features/attributes that are sent to CRFsuite - they could be helpful e.g. to check which attributes (returned by info()) are active for a given observation.

Initialize ItemSequence with a list of item features:

>>> ItemSequence([{'foo': 1, 'bar': 0}, {'foo': 1.5, 'baz': 2}])
<ItemSequence of size 2>

Item features could be in one of the following formats:

  • {“string_key”: float_weight, …} dict where keys are observed features and values are their weights;
  • {“string_key”: bool, …} dict; True is converted to 1.0 weight, False - to 0.0;
  • {“string_key”: “string_value”, …} dict; that’s the same as {“string_key=string_value”: 1.0, …}
  • [“string_key1”, “string_key2”, …] list; that’s the same as {“string_key1”: 1.0, “string_key2”: 1.0, …}
  • {“string_prefix”: {…}} dicts: nested dict is processed and “string_prefix” s prepended to each key.
  • {“string_prefix”: […]} dicts: nested list is processed and “string_prefix” s prepended to each key.
  • {“string_prefix”: set([…])} dicts: nested list is processed and “string_prefix” s prepended to each key.

Dict-based features can be mixed, i.e. this is allowed:

{"key1": float_weight,
 "key2": "string_value",
 "key3": bool_value,
 "key4: {"key5": ["x", "y"], "key6": float_value},
 }
items(self)

Return a list of prepared item features: a list of {unicode_key: float_value} dicts.

>>> ItemSequence([["foo"], {"bar": {"baz": 1}}]).items()
[{'foo': 1.0}, {'bar:baz': 1.0}]

Training

class pycrfsuite.Trainer

Bases: pycrfsuite._pycrfsuite.BaseTrainer

The trainer class.

This class maintains a data set for training, and provides an interface to various training algorithms.

Parameters:
algorithm : {‘lbfgs’, ‘l2sgd’, ‘ap’, ‘pa’, ‘arow’}

The name of the training algorithm. See Trainer.select().

params : dict, optional

Training parameters. See Trainer.set_params() and Trainer.set().

verbose : boolean

Whether to print debug messages during training. Default is True.

append(self, xseq, yseq, int group=0)

Append an instance (item/label sequence) to the data set.

Parameters:
xseq : a sequence of item features

The item sequence of the instance. xseq should be a list of item features or an ItemSequence instance. Allowed item features formats are the same as described in ItemSequence docs.

yseq : a sequence of strings

The label sequence of the instance. The number of elements in yseq must be identical to that in xseq.

group : int, optional

The group number of the instance. Group numbers are used to select subset of data for heldout evaluation.

clear(self)

Remove all instances in the data set.

get(self, name)

Get the value of a training parameter. This function gets a parameter value for the graphical model and training algorithm specified by Trainer.select() method.

Parameters:
name : string

The parameter name.

get_params(self)

Get training parameters.

Returns:
dict

A dictionary with {parameter_name: parameter_value} with all trainer parameters.

help(self, name)

Get the description of a training parameter. This function obtains the help message for the parameter specified by the name. The graphical model and training algorithm must be selected by Trainer.select() method before calling this method.

Parameters:
name : string

The parameter name.

Returns:
string

The description (help message) of the parameter.

logparser = None
message(self, message)
on_end(self, log)
on_featgen_end(self, log)
on_featgen_progress(self, log, percent)
on_iteration(self, log, info)
on_optimization_end(self, log)
on_prepare_error(self, log)
on_prepared(self, log)
on_start(self, log)
params(self)

Obtain the list of parameters.

This function returns the list of parameter names available for the graphical model and training algorithm specified in Trainer constructor or by Trainer.select() method.

Returns:
list of strings

The list of parameters available for the current graphical model and training algorithm.

select(self, algorithm, type='crf1d')

Initialize the training algorithm.

Parameters:
algorithm : {‘lbfgs’, ‘l2sgd’, ‘ap’, ‘pa’, ‘arow’}

The name of the training algorithm.

  • ‘lbfgs’ for Gradient descent using the L-BFGS method,
  • ‘l2sgd’ for Stochastic Gradient Descent with L2 regularization term
  • ‘ap’ for Averaged Perceptron
  • ‘pa’ for Passive Aggressive
  • ‘arow’ for Adaptive Regularization Of Weight Vector
type : string, optional

The name of the graphical model.

set(self, name, value)

Set a training parameter. This function sets a parameter value for the graphical model and training algorithm specified by Trainer.select() method.

Parameters:
name : string

The parameter name.

value : string

The value of the parameter.

set_params(self, params)

Set training parameters.

Parameters:
params : dict

A dict with parameters {name: value}

train(self, model, int holdout=-1)

Run the training algorithm. This function starts the training algorithm with the data set given by Trainer.append() method.

Parameters:
model : string

The filename to which the trained model is stored. If this value is empty, this function does not write out a model file.

holdout : int, optional

The group number of holdout evaluation. The instances with this group number will not be used for training, but for holdout evaluation. Default value is -1, meaning “use all instances for training”.

verbose

verbose: object

Tagging

class pycrfsuite.Tagger

The tagger class.

This class provides the functionality for predicting label sequences for input sequences using a model.

close(self)

Close the model.

dump(self, filename=None)

Dump a CRF model in plain-text format.

Parameters:
filename : string, optional

File name to dump the model to. If None, the model is dumped to stdout.

info(self)

Return a ParsedDump structure with model internal information.

labels(self)

Obtain the list of labels.

Returns:
list of strings

The list of labels in the model.

marginal(self, y, pos)

Compute the marginal probability of the label y at position pos for the current input sequence (i.e. a sequence set using Tagger.set() method or a sequence used in a previous Tagger.tag() call).

Parameters:
y : string

The label.

t : int

The position of the label.

Returns:
float

The marginal probability of the label y at position t.

open(self, name)

Open a model file.

Parameters:
name : string

The file name of the model file.

open_inmemory(self, bytes value)

Open a model from memory.

Parameters:
value : bytes

Binary model data (content of a file saved by Trainer.train).

probability(self, yseq)

Compute the probability of the label sequence for the current input sequence (a sequence set using Tagger.set() method or a sequence used in a previous Tagger.tag() call).

Parameters:
yseq : list of strings

The label sequence.

Returns:
float

The probability P(yseq|xseq).

set(self, xseq)

Set an instance (item sequence) for future calls of Tagger.tag(), Tagger.probability() and Tagger.marginal() methods.

Parameters:
xseq : item sequence

The item sequence of the instance. xseq should be a list of item features or an ItemSequence instance. Allowed item features formats are the same as described in ItemSequence docs.

tag(self, xseq=None)

Predict the label sequence for the item sequence.

Parameters:
xseq : item sequence, optional

The item sequence. If omitted, the current sequence is used (a sequence set using Tagger.set() method or a sequence used in a previous Tagger.tag() call).

xseq should be a list of item features or an ItemSequence instance. Allowed item features formats are the same as described in ItemSequence docs.

Returns:
list of strings

The label sequence predicted.

Debugging

class pycrfsuite._dumpparser.ParsedDump[source]

CRFsuite model parameters. Objects of this type are returned by pycrfsuite.Tagger.info() method.

Attributes:
transitions : dict

{(from_label, to_label): weight} dict with learned transition weights

state_features : dict

{(attribute, label): weight} dict with learned (attribute, label) weights

header : dict

Metadata from the file header

labels : dict

{name: internal_id} dict with model labels

attributes : dict

{name: internal_id} dict with known attributes