class model_trainer:
def __init__(self):
# read model/transformer from file
self.model_file_dir = "core_model"
self.count_vect_file_name = "core_model/count_vectorizer.model"
if not os.path.isfile(self.count_vect_file_name):
print "Build transformer first! (python build_text_transformer.py)"
sys.exit()
self.count_vect = joblib.load(self.count_vect_file_name)
self.prep = Preprocessor()
self.model_major = [
('SVM',
BaggingClassifier(SGDClassifier(loss='hinge',
penalty='l2',
alpha=1e-3,
n_iter=5,
random_state=42),
max_samples=0.5,
max_features=0.5)),
('NB',
BaggingClassifier(MultinomialNB(alpha=.01),
max_samples=0.5,
max_features=0.5)),
# ('ANN' , BaggingClassifier(MLPClassifier(solver='lbfgs', alpha=1e-5,hidden_layer_sizes=(5, 2), random_state=1),max_samples=0.5, max_features=0.5)),
# ('KNN' , BaggingClassifier(KNeighborsClassifier(n_neighbors=10),max_samples=0.5, max_features=0.5)),
('RDFOREST', RandomForestClassifier(n_estimators=25)),
('NC',
BaggingClassifier(NearestCentroid(),
max_samples=0.5,
max_features=0.5)),
('ADA-SAMME.R', AdaBoostClassifier(n_estimators=100)),
]
self.models = {
'SVM':
SGDClassifier(loss='hinge',
penalty='l2',
alpha=1e-3,
n_iter=5,
random_state=42),
'NB':
MultinomialNB(alpha=.01),
# 'ANN' : MLPClassifier(solver='lbfgs', alpha=1e-5,hidden_layer_sizes=(5, 2), random_state=1),
# 'KNN' : KNeighborsClassifier(n_neighbors=10),
'RDFOREST':
RandomForestClassifier(n_estimators=25),
'NC':
NearestCentroid(),
# 'MAJOR' : VotingClassifier(estimators=self.model_major,voting='soft',n_jobs=-1)
}
def load_data(self, size=0):
self.prep.load(size)
def train(self, X, y, count_vect, clf, partial=False):
X_count = count_vect.transform(X)
X_tfidf = TfidfTransformer().fit_transform(X_count)
if partial:
clf.partial_fit(X_tfidf, y)
else:
clf.fit(X_tfidf, y)
def predict(self, X, count_vect, clf):
X_count = count_vect.transform(X)
X_tfidf = TfidfTransformer().fit_transform(X_count)
return clf.predict(X_tfidf)
def train_all_tag(self):
train_tag_list = []
heightest_score = defaultdict(float)
heightest_info = {}
heightest_model = {}
all_tag_idx = self.prep.get_all_tag_idx()
for target_tag in all_tag_idx:
for model_name in self.models:
X_train, y_train, X_test, y_test = self.prep.get_train_test_data_tag(
target_tag)
if len(y_train) < 200:
print "[Train] not enough (%3d less than 100) sample for '%s'" % (
len(y_train), self.prep.get_target_names()
[target_tag].encode('utf-8'))
continue
# use only content from (paragraph_id,content)
X_train = [data[1] for data in X_train]
X_test = [data[1] for data in X_test]
train_tag_list.append(target_tag)
print "train %s" % target_tag
self.train(X_train, y_train, self.count_vect,
self.models[model_name], False)
predicted = self.predict(X_test, self.count_vect,
self.models[model_name])
score = np.mean(predicted == y_test)
matrix = metrics.precision_recall_fscore_support(
y_test, predicted, average='binary', pos_label=target_tag)
precision = matrix[0]
recall = matrix[1]
f1 = matrix[2]
print "%s score %.2f (%s)-[%d/%d]" % (
model_name, score,
self.prep.get_target_names()[target_tag].encode('utf-8'),
len(y_train), len(y_test))
print "precision=%.2f, recall=%.2f, *** f1=%.2f ***" % (
precision, recall, f1)
print
if heightest_score[target_tag] < f1:
heightest_score[target_tag] = f1
heightest_info[target_tag] = (f1, precision, recall, score,
model_name)
heightest_model[target_tag] = deepcopy(
self.models[model_name])
db = DB()
db.clear_model_score()
for tag in set(train_tag_list):
if not tag in heightest_info:
continue
output_filename = os.path.join(self.model_file_dir,
"%s.model" % tag)
print "[Train] save model '%s' (%s) with F1=%.2f, precision=%.2f, recall=%.2f, score=%.2f : %s" % (
output_filename,
self.prep.get_target_names()[tag].encode('utf-8'),
heightest_info[tag][0], heightest_info[tag][1],
heightest_info[tag][2], heightest_info[tag][3],
heightest_info[tag][4])
joblib.dump(heightest_model[tag], output_filename)
db.add_model_info(tag, output_filename, heightest_score[tag])
return heightest_info
