def _liblinear_classify(vecs, model):
from linearutil import predict as liblinear_predict
import sys
orig_stdout = sys.stdout
try:
with open('/dev/null', 'w') as dev_null:
sys.stdout = dev_null
# We don't really need -b 1 here
predictions, _, _ = liblinear_predict(
[], vecs, model, '') #'-b 1')
return predictions
finally:
sys.stdout = orig_stdout
def _liblinear_classify(self, vec, ranked=False):
import sys
orig_stdout = sys.stdout
try:
with open('/dev/null', 'w') as dev_null:
sys.stdout = dev_null
# Ask for probs. when it is necessary
if not ranked:
args = ''
else:
args = '-b 1'
from linearutil import predict as liblinear_predict
predictions, _, probabilities = liblinear_predict(
[], [vec], self.model, args)
finally:
sys.stdout = orig_stdout
try:
if not ranked:
return self.name_by_lbl_id[predictions[0]]
else:
# We need a bit of magic to get this list right since LibLinear
# only returns a single label we will played with the indexes
# NOTE: Labels are stored in self.model.label, never assume otherwise
probs_and_lbl = [(prob, self.name_by_lbl_id[lbl_id])
for lbl_id, prob in izip(self.model.label, probabilities[0])]
probs_and_lbl.sort()
probs_and_lbl.reverse()
# Now flip the tuples and we have the ranked labels
return [(lbl, prob) for prob, lbl in probs_and_lbl]
except KeyError:
print predictions
print self.name_by_lbl_id
raise
