def predict(self, X):
"""
Uses trained support vector classifier to predict label.
"""
f_name = os.path.join(abs_path("../model"), self.name + '.predict')
first = True
with open(f_name, 'w') as f:
for x in X:
out = ("|BT| " + str(x.getEnrichedTree(self.mode)) + " |ET|"
+ hotty(x.getConcepts()[0].label, x.getConcepts()[1].label)
+ " |EV|\n")
f.write(out)
for label in _LABELS.values():
m_name = os.path.join(abs_path("../model"), self.name + '.' + label + '.svm')
r_name = os.path.join(abs_path("../results"), self.name + '.' + label + '.predict')
subprocess.call([abs_path('../bin/svm-light-TK-1.2.1/svm_classify'),
f_name,
m_name,
r_name])
r_label = list()
with open(r_name, 'r') as f:
for line in f:
r_label.append(float(line))
if first:
results = np.array(r_label).reshape(len(r_label),1)
first = False
else:
results = np.concatenate((results,np.array(r_label).reshape(len(r_label),1)),axis=1)
os.remove(r_name)
os.remove(f_name)
return [_LABELS[r] for r in results.argmax(axis=1)]
def train(self, X, Y):
"""
Trains a classifier using a custom kernel.
Note: X is a list of entries Y is a list of labels
"""
for label in _LABELS.values():
f_name = os.path.join(abs_path('../model'), self.name + '.' + label + '.tmp')
with open(f_name, 'w') as f:
for y, x in zip(Y,X):
if y == label:
y = '1'
else:
y = '-1'
out = (str(y) + " |BT| " + str(x.getEnrichedTree(self.mode)) + " |ET|"
+ hotty(x.getConcepts()[0].label, x.getConcepts()[1].label)
+ " |EV|\n")
f.write(out)
subprocess.call([abs_path('../bin/svm-light-TK-1.2.1/svm_learn'),
'-t', '5',
'-S', '1',
'-D', '1',
'-r', '1',
'-c', '0.5',
'-d', '2',
'-T', '3.35',
'-N', '3',
'-W', 'S',
'-v', '0',
f_name,
f_name.split('.tmp')[0] + '.svm',
])
# Create Model
if (v):
sys.stdout.write("\tCreating model...\n")
if model_type == 'svm-spt':
svm.MODE = 'spt'
ml = svm
elif model_type == 'svm-insert':
svm.MODE = 'insert'
ml = svm
elif model_type == 'svm-suffix':
svm.MODE = 'suffix'
ml = svm
else:
sys.stderr.write("ERROR: Invalid model type.\n")
sys.exit(1)
model = ml.Model()
model.train(entries, labels)
# Save model for later use
if (v):
sys.stdout.write("\tPickling model to %s...\n" % model_path)
model.save(model_path)
if __name__ == "__main__":
main(abs_path('./i2b2_examples/'), abs_path('../model/example-spt.mod'), 'svm-spt', True)
main(abs_path('./i2b2_examples/'), abs_path('../model/example-insert.mod'), 'svm-insert', True)
main(abs_path('./i2b2_examples/'), abs_path('../model/example-suffix.mod'), 'svm-suffix', True)
total_p_labels += p_labels
total_g_labels += g_labels
# Calculate Precision and Recall Using true positive, false positive, false
# negative for relation detection
precision = float(total_TP) / (total_TP + total_FP)
recall = float(total_TP) / (total_TP + total_FN)
F1 = 2 * (precision * recall) / (precision + recall)
with open(results_file, "w") as f:
print >>f, "-" * 80
print >>f, "Relation Detection Statistics:"
print >>f, ""
print >>f, "\tTrue positives: ", total_TP
print >>f, "\tFalse positives:", total_FP
print >>f, "\tFalse negatives:", total_FN
print >>f, ""
print >>f, "\tPrecision: %.4f \t Recall: %.4f \t F1: %.4f" % (precision, recall, F1)
print >>f, ""
print >>f, "-" * 80
print >>f, "Classification Statistics:"
print >>f, ""
print >>f, classification_report(total_g_labels, total_p_labels)
print >>f, "-" * 80
print >>f, ""
if __name__ == "__main__":
main(abs_path("../results/"),
abs_path("./i2b2_examples/rel/"),
abs_path("../results/i2b2_results"), True)