class NBClassifier(TransformerMixin):
"""Naive Bayes classifier for part-of-text classification.
The classifier creates a wrapper around NLTK NaiveBayesClassifier
and implements `transform` and `fit_transform` methods suitable for
pipeline integration.
:param label_probdist: P(label)
The probability distribution over labels.
It is expressed as a ``ProbDistI`` whose samples are labels.
I.e., P(label) = ``label_probdist.prob(label)``.
:param feature_probdist: P(fname=fval|label)
The probability distribution for feature values, given labels.
It is expressed as a dictionary whose keys are ``(label, fname)``
pairs and whose values are ``ProbDistI`` objects over feature values.
I.e., P(fname=fval|label) = ``feature_probdist[label,fname].prob(fval)``.
If a given ``(label,fname)`` is not a key in ``feature_probdist``,
then it is assumed that the corresponding P(fname=fval|label)
is 0 for all values of ``fval``.
"""
def __init__(self,
label_probdist=None,
feature_probdist=None,
estimator=ELEProbDist):
"""Initialize NBClassifier."""
self._estimator = estimator
# in case arguments are specified (ie. when restoring the classifier)
if all([label_probdist, feature_probdist]):
self._classifier = NaiveBayesClassifier(
label_probdist=label_probdist,
feature_probdist=feature_probdist,
)
else:
self._classifier = None
@property
def features(self):
"""Return features most informative for classification."""
if self._classifier is None:
return None
return self._classifier.most_informative_features()
# noinspection PyPep8Naming, PyUnusedLocal
def fit(
self,
X: typing.Iterable, # pylint: disable=invalid-name
y=None, # pylint: disable=unused-argument
**fit_params):
"""Fits the classifier to the given data set.
:param X: Iterable, output of FeatureExtractor
The X is expected to be an iterable of tuples (tagged_word, feature_set, label),
where feature set is a dictionary of evaluated features.
The format of X matches the output of `FeatureExtractor`.
:param y: redundant (included to preserve base class method definition)
"""
# NLTK classifier expects stacked featuresets for the training,
# so we need to reduce the dimenstionality
labeled_featuresets = list()
for entry in X:
labeled_featuresets.extend([
(featureset, feature_label)
for _, featureset, feature_label in entry
])
# initialize the NLTK classifier
self._classifier = NaiveBayesClassifier.train(
labeled_featuresets, estimator=self._estimator)
return self
# noinspection PyPep8Naming, PyUnusedLocal
def transform(self, X): # pylint: disable=invalid-name,unused-argument
"""Auxiliary function to be used in pipeline."""
return self
# noinspection PyPep8Naming
def evaluate(
self,
X: typing.Iterable, # pylint: disable=invalid-name
y: typing.Iterable,
sample,
n=3,
filter_hooks=None):
"""Perform evaluation of the classifier instance.
:param X: Iterable, test data
Same shape as for `fit` and `fit_predict` methods
:param y: Iterable, of labels
:param sample: one of labels to get the prediction for
For example, if labels are ['class_A', 'class_B', 'class_C'], the sample
could be 'class_A'.
:param n: int, number of candidates to output
:param filter_hooks: list of hooks, will be used to filter predictions
The hook should take a tuple of ((word, tag), score) as its parameter
and output boolean whether or not it passes the filter.
"""
# noinspection PyTypeChecker,PyTypeChecker
if len(X) != len(y):
raise ValueError("`X` and `y` must be of the same length.")
candidate_arr = self.fit_predict(X,
n=n,
sample=sample,
filter_hooks=filter_hooks or [])
correctly_predicted = 0
for candidates, label in zip(candidate_arr, y):
pred = self._valid_candidates(candidates, label)
correctly_predicted += int(pred)
# return the accuracy score
# noinspection PyTypeChecker
return precision(total=len(y), correct=correctly_predicted)
# noinspection PyPep8Naming
def fit_predict(
self,
X: typing.Iterable, # pylint: disable=invalid-name
y=None, # pylint: disable=unused-argument
**fit_params):
"""Make prediction about the given data.
:param X: Iterable, prediction data
The prediction data is expected to be of type
List[(name_tuple, feature_set [,feature,label)] where feature_set
corresponds to the output of FeatureExtractor and feature
labels (if provided) should be None (will be ignored anyway).
:param y: redundant (included to preserve bace class method definition)
:param fit_params: kwargs, fit parameters
n: number of candidates to output
sample: one of labels to get the prediction for (for example,
if labels are ['class_A', 'class_B', 'class_C'], the sample
could be 'class_A'.
filter_hooks: list of hooks, will be used to filter predictions
The hook should take a tuple of ((word, tag), score) as its parameter
and output boolean whether or not it passes the filter.
"""
# get fit parameters
n = fit_params.get('n', 3)
sample = fit_params.get('sample', None)
# do not allow sample to be `None` (wouldn't be possible to sort
# the candidates in a logical way)
if sample is None:
raise ValueError("`fit_parameter` `sample` was not specified."
" This is not allowed in `fit_predict` method")
if not all([hasattr(var, '__len__') for var in [X, y or []]]):
raise TypeError("`X` and `y` must implement `__len__` method")
# noinspection PyTypeChecker
predictions = [None] * len(X)
for i, x in enumerate(X):
candidate_pred = [None] * len(x)
for j, candidate in enumerate(x):
if len(candidate) == 3:
# feature label was provided as part of X set (usual case), ignore it
name_tuple, features, _ = candidate
else:
name_tuple, features = candidate
candidate_pred[j] = (name_tuple,
self.predict(features, sample=sample))
sorted_pred = sorted(candidate_pred,
key=lambda t: t[1],
reverse=True)
for hook in fit_params.get('filter_hooks', []):
sorted_pred = list(filter(hook, sorted_pred))
predictions[i] = sorted_pred[:n]
return np.array(predictions)
def predict(self, features: dict, sample=None) -> typing.Any:
"""Make predictions based on given features.
:param features: dict, features to be used for prediction
Dictionary of (feature_key, feature_value)
:param sample:
one of labels to get the prediction for (for example,
if labels are ['class_A', 'class_B', 'class_C'], the sample
could be 'class_A'.
:returns: Union[float, dict]
If `sample` is specified, returns P(sample|features),
ie the probability of `sample` given features,
where `sample` is one of labels.
Otherwise returns dict of (label: max_prob) for all
known labels.
"""
if self._classifier is None:
raise ValueError("Unable to make predictions. "
"Classifier has not been trained yet!")
prob_dist = self._classifier.prob_classify(features)
# sort by the probability
if sample is not None:
probs = prob_dist.prob(sample)
else:
probs = {s: prob_dist.prob(s) for s in self._classifier.labels()}
return probs
def show_most_informative_features(self):
"""Print features most informative for classification."""
if self._classifier is None:
return
self._classifier.show_most_informative_features()
def export(self, export_dir=None, export_name=None) -> str:
"""Export timestamped pickled classifier to the given directory.
:returns: path to the timestamped .checkpoint file
"""
export_dir = export_dir or 'export/'
export_name = export_name or 'classifier'
if export_name.endswith('.checkpoint'):
export_name = ".".join(export_name.split('.')[:-1])
time_stamp = str(datetime.datetime.now().timestamp())
# create export directory
os.makedirs(export_dir, exist_ok=True)
time_stamped_fname = ".".join([export_name, time_stamp, 'checkpoint'])
time_stamped_fpath = os.path.join(export_dir, time_stamped_fname)
# pickle and export the classifier
with open(time_stamped_fpath, 'wb') as exp_file:
pickle.dump(self, exp_file)
return time_stamped_fpath
@staticmethod
def restore(checkpoint) -> "NBClassifier":
"""Restores the classifier from a checkpoint file.
:param checkpoint: path to directory or specific checkpoint
If path to directory provided, the newest checkpoint
is restored.
"""
def _restore_checkpoint(fp):
with open(fp, 'rb') as checkpoint_file:
# load the exported classifier
return pickle.load(checkpoint_file)
if os.path.isdir(checkpoint):
checkpoint_dir = checkpoint
checkpoints = [
os.path.join(checkpoint_dir, f) for f in os.listdir(checkpoint)
if f.endswith('.checkpoint')
]
# find the latest
if not checkpoints:
raise ValueError(
"No checkpoints were found in `{}`.".format(checkpoint))
latest_checkpoint = sorted(checkpoints)[-1]
clf = _restore_checkpoint(latest_checkpoint)
else:
clf = _restore_checkpoint(checkpoint)
return clf
@staticmethod
def _valid_candidates(candidates: typing.Iterable, label):
"""Check whether the correct label is among candidates."""
for candidate, _ in candidates:
# FIXME: a bug here, NLTK lets weird things like '**' go through -> causes crash
candidate_name, _ = candidate
try:
if re.search(candidate_name, label, flags=re.IGNORECASE):
return True
except Exception:
return False
return False
