class TranslationClassifier(object):
"""
Class for translation disambiguation using per lempos disambiguators
stored in a HDF5 file
"""
def __init__(self, models_fname):
self.models = DisambiguatorStore(models_fname)
self.classifier = self.models.load_estimator()
self.vocab = self.models.load_vocab(as_dict=True)
def score(self, source_lempos, context_vec):
"""
score translation candidates for source lempos combination,
returning a dict mapping target lempos combinations to scores
"""
try:
self.models.restore_fit(source_lempos, self.classifier)
except KeyError:
log.debug(u"no model available for source lempos " +
source_lempos)
return {}
target_names = self.models.load_target_names(source_lempos)
return self._predict(context_vec, target_names)
def _predict(self, context_vec, target_names):
"""
return a dict mapping target names to scores
"""
# FIXME: some estimators have no predict_proba method
# e.g. NearestCentroid
preds = self.classifier.predict_proba(context_vec)
return dict(zip(target_names, preds[0]))
def _prepare(self):
self.start_time = time.time()
self.disambiguator_count = 0
log.info("creating models file " + self.models_hdf_fname)
self.models_hdfile = DisambiguatorStore(self.models_hdf_fname, "w")
self.models_hdfile.save_estimator(self.classifier)
# FIXME: hmm, a bit sneaky...
self.models_hdfile.copy_vocab(self.data_generator.samp_hdfile)
if self.counts_fname:
log.info("reading counts from " + self.counts_fname)
self.counts_dict = cPickle.load(open(self.counts_fname))
if self.with_vocab_mask:
log.info("storage with vocabulary masks")
self.vocab = self.models_hdfile.load_vocab()
def test_disambiguator_store(self):
# Create a silly classifier that disambiguates between "stam" (tree
# trunk) or "romp" (body trunk) as the Dutch translation of the
# English noun "trunk"
lempos = u"trunk/n"
# FIXME: store_fit() should only accept unicode strings
target_names = u"stam romp".encode("utf-8").split()
vocab = u"boom hoofd".split()
X = np.array([[0,1],
[1,0],
[0,1],
[1,0]])
y = np.array([1,0,1,0])
estimator = NearestCentroid()
estimator.fit(X, y)
centroids = estimator.centroids_
score = estimator.score(X, y)
# Store estimator
fname = tempfile.NamedTemporaryFile().name
f = DisambiguatorStore(fname, "w")
f.save_estimator(NearestCentroid())
f.save_vocab(vocab)
f.store_fit(lempos, estimator)
f.save_target_names(lempos, target_names)
f.close()
# Restore estimator
f2 = DisambiguatorStore(fname)
estimator2 = f2.load_estimator()
vocab2 = f2.load_vocab()
f2.restore_fit(lempos, estimator2)
target_names2 = f2.load_target_names(lempos)
centroids2 = estimator2.centroids_
score2 = estimator2.score(X, y)
assert_array_equal(centroids, centroids2)
assert target_names == target_names2
assert vocab == vocab2
assert score == score2
def trace_nc(graphs_fname, model_fname, n=None, source_pos=[],
outf=codecs.getwriter('utf8')(sys.stdout)):
if isinstance(outf, basestring):
outf = codecs.open(outf, "w", encoding="utf-8")
graphs = cPickle.load(open(graphs_fname))
model = DisambiguatorStore(model_fname)
estimator = model.load_estimator()
vocab = np.array(model.load_vocab())
vocab_dict = model.load_vocab(as_dict=True)
score_attr = "centroid_score"
for graph_count, graph in enumerate(graphs[:n]):
source_string = graph.source_string()
outf.write( 100 * "=" + "/n")
outf.write( u"{}: {}\n".format(graph_count + 1, source_string))
outf.write( 100 * "=" + "\n\n")
reverse_lookup = make_reverse_lookup(graph)
source_node_vectors = make_source_node_vectors(graph, vocab_dict)
source_graph_vector = sp.csr_matrix(source_node_vectors.sum(axis=0))
for sn, node_vec in zip(graph.source_nodes_iter(ordered=True),
source_node_vectors):
if source_pos and graph.pos(sn) not in source_pos:
continue
try:
source_lempos = u" ".join(graph.node[sn]["lex_lempos"])
except KeyError:
log.debug(u"(annotated) source lempos {0} not in"
u"lexicon\n".format(graph.lempos(sn)))
continue
try:
model.restore_fit(source_lempos, estimator)
except KeyError:
log.debug(u"no model available for (lexicon) source lempos "
u"{0}\n".format(source_lempos))
continue
context_vec = source_graph_vector - node_vec
context_vec = context_vec.toarray()
normalize(context_vec, copy=False)
try:
mask = model.load_vocab_mask(source_lempos)[:]
except KeyError:
local_vocab = vocab
else:
context_vec = context_vec[:,mask]
local_vocab = vocab[mask]
try:
centroids = estimator.centroids_
except AttributeError:
# pipeline
centroids = estimator.steps[-1][-1].centroids_
target_lempos_list = model.load_target_names(source_lempos)
outf.write( 100 * "-" + "\n")
outf.write( source_lempos + "\n")
outf.write( 100 * "-" + "\n\n")
for target_lempos, target_centroid in zip(target_lempos_list,
centroids):
prod = target_centroid * context_vec
outf.write( u"==> {:<24} {:1.4f} {}\n".format(
target_lempos,
prod.sum(),
prod.sum() * 100 * "X"))
outf.write("\n")
for target_lempos, target_centroid in zip(target_lempos_list,
centroids):
prod = target_centroid * context_vec
indices = target_centroid.argsort()[::-1]
outf.write( "\n" + source_string + "\n\n")
outf.write( u"{:<64} ==> {:<24} {:1.4f} {}\n".format(
source_lempos,
target_lempos,
prod.sum(),
prod.sum() * 100 * "X"))
for i in indices:
if prod[0,i] > 0:
context_lemma = local_vocab[i]
sources = ",".join(reverse_lookup[context_lemma])
bar = prod[0,i] * 100 * "*"
outf.write( u"{:<64} --> {:<24} {:1.4f} {}\n".format(
sources,
context_lemma,
prod[0,i],
bar))
outf.write("\n")
class ModelBuilder(object):
"""
Class for fitting disambiguators on samples and storing them in HDF5 file.
Parameters
----------
data_generator: DataSetGenerator instance
Labeled samples generator
samp_hdf_fname: str
Name of HDF5 file containing context samples
models_hdf_fname: str
Name of HDF5 file for storing disambiguation models
classifier: classifier instance
Classifier instance from sklearn, to be fitted for each ambiguous
source lempos. Possibly a Pipeline instance.
counts_fname: str
Name of pickle file containing lemma counts. If used in combination
with classifier that supports class weights (e.g. SGDClassifier)
or class priors (e.g. Naive Bayes), classes will be weighted
according to their lemma frequency.
with_vocab_mask: bool, optional
If true, a vocabulary mask is stored for each model. This is required
when the classifier performs feature selection and the vocabulary has
to be pruned accordingly.
"""
# feature selectors that reduce the vocabulary
FEATURE_SELECTORS = _BaseFilter, RFE, _LearntSelectorMixin
def __init__(self, data_generator, models_hdf_fname, classifier,
counts_fname=None, with_vocab_mask=False, **kwargs):
self.data_generator = data_generator
self.models_hdf_fname = models_hdf_fname
self.classifier = classifier
self.counts_fname = counts_fname
self.with_vocab_mask = with_vocab_mask
def run(self):
"""
build disambiguation models
"""
self._prepare()
self._build()
self._finish()
def _prepare(self):
self.start_time = time.time()
self.disambiguator_count = 0
log.info("creating models file " + self.models_hdf_fname)
self.models_hdfile = DisambiguatorStore(self.models_hdf_fname, "w")
self.models_hdfile.save_estimator(self.classifier)
# FIXME: hmm, a bit sneaky...
self.models_hdfile.copy_vocab(self.data_generator.samp_hdfile)
if self.counts_fname:
log.info("reading counts from " + self.counts_fname)
self.counts_dict = cPickle.load(open(self.counts_fname))
if self.with_vocab_mask:
log.info("storage with vocabulary masks")
self.vocab = self.models_hdfile.load_vocab()
def _build(self):
for data_set in self.data_generator:
if data_set.target_lempos:
self._build_disambiguator(data_set)
else:
log.error("no samples and thus no disambiguation models for " +
data_set.source_lempos)
def _build_disambiguator(self, data_set):
log.info(u"building disambiguator for {} with {} translations".format(
data_set.source_lempos, len(data_set.target_lempos)))
if self.counts_fname:
self._set_class_weights(data_set.target_lempos)
try:
self.classifier.fit(data_set.samples, data_set.targets)
except ValueError, error:
if ( error.args in [
("zero-size array to reduction operation maximum which has no "
"identity",),
("zero-size array to maximum.reduce without identity",),
('Invalid threshold: all features are discarded.',)] ):
# this happens when there are no features selected
# e.g. when using SelectFpr
log.error("No model created, because no features selected!")
return
else:
raise error
self.models_hdfile.store_fit(data_set.source_lempos, self.classifier)
self.models_hdfile.save_target_names(data_set.source_lempos,
data_set.target_lempos)
if self.with_vocab_mask:
self._save_vocab_mask(data_set.source_lempos)
self.disambiguator_count += 1
def print_centroids(models_fname, lemma=None, pos=None, minimum=0, n=None,
outf=codecs.getwriter('utf8')(sys.stdout) ):
# If used in combination with a feature selector,
# models must be build using with_vocab_mask=True
# FIXME: messy code below
models = DisambiguatorStore(models_fname)
classifier = models.load_estimator()
full_vocab = np.array(models.load_vocab())
fits = models.file[models.FITS_PATH]
line = 78 * "=" + "\n"
subline = " " + 74 * "-" + "\n"
if isinstance(outf, basestring):
outf = codecs.open(outf, "w", encoding="utf-8")
if lemma:
if lemma in fits:
lemma_list = [lemma]
else:
lemma_list = []
else:
lemma_list = fits
for lemma in lemma_list:
for lemma_pos in fits[lemma]:
if not pos or lemma_pos == pos:
lempos = lemma + u"/" + lemma_pos
outf.write(line + lempos + "\n" + line)
models.restore_fit(lempos, classifier)
if isinstance(classifier, NearestCentroid):
centroids_ = classifier.centroids_
else:
# NearestCentroid is last item in Pipeline
nc = classifier.steps[-1][-1]
assert isinstance(nc, NearestCentroid)
centroids_ = nc.centroids_
target_names = models.load_target_names(lempos)
target_n = 0
try:
vocab_mask = models.load_vocab_mask(lempos)[:]
except:
vocab = full_vocab
else:
vocab = full_vocab[vocab_mask]
for target, centroid in zip(target_names, centroids_):
target_n += 1
outf.write(subline)
outf.write(u" [{}] {} ---> {}\n".format(
target_n,
lempos,
target))
outf.write(subline)
indices = centroid.argsort().tolist()
indices.reverse()
for i in indices[:n]:
if centroid[i] > minimum:
outf.write(u" {0:>16.8f} {1:<16} {2}\n".format(
centroid[i],
vocab[i],
centroid[i] * 100 * "*"))
def __init__(self, models_fname):
self.models = DisambiguatorStore(models_fname)
self.classifier = self.models.load_estimator()
self.vocab = self.models.load_vocab(as_dict=True)
