def predict(self, features, level):
if level == 0:
vec = vectorize(features, self.feature_template_level_0)
pred_label = self.classifier_level_0.predict(vec)
if level == 1:
vec = vectorize(features, self.feature_template_level_1)
pred_label = self.classifier_level_1.predict(vec)
if level == 2:
vec = vectorize(features, self.feature_template_level_2)
pred_label = self.classifier_level_2.predict(vec)
return self.idxrelation_map[pred_label[0]]
def gen_relation_train_data(self, level):
for rst_tree in self.rst_tree_instances:
for feats, relation in rst_tree.generate_relation_samples(
self.brown_clusters, level):
if level == 0:
yield vectorize(feats, self.relation_feat_template_level_0
), self.relation_map[relation]
if level == 1:
yield vectorize(feats, self.relation_feat_template_level_1
), self.relation_map[relation]
if level == 2:
yield vectorize(feats, self.relation_feat_template_level_2
), self.relation_map[relation]
def predict(self, features, dependency_feat, level):
if level == 0:
vec = vectorize(features, self.feature_template_level_0)
# pred_label = self.classifier_level_0.predict(vec)
pred_label = self.classifier_level_0.predict(np.concatenate([vec.toarray(), dependency_feat], axis=1))
if level == 1:
vec = vectorize(features, self.feature_template_level_1)
# pred_label = self.classifier_level_1.predict(vec)
pred_label = self.classifier_level_1.predict(np.concatenate([vec.toarray(), dependency_feat], axis=1))
if level == 2:
vec = vectorize(features, self.feature_template_level_2)
# pred_label = self.classifier_level_2.predict(vec)
pred_label = self.classifier_level_2.predict(np.concatenate([vec.toarray(), dependency_feat], axis=1))
return self.idxrelation_map[pred_label[0]]
def gen_action_train_data(self):
tree = []
for rst_tree in self.rst_tree_instances:
print("gen_action_train_data ...")
for feats, dependency_feats, action in rst_tree.generate_action_samples(
self.brown_clusters):
yield vectorize(feats, self.action_feat_template
), self.action_map[action], dependency_feats[0]
tree.append(rst_tree)
def gen_relation_train_data(self, level):
# trees = pickle.load(open('../data/model/action/tree.bin','rb'))
for rst_tree in self.rst_tree_instances:
print("gen_relation_train_data ...")
# for rst_tree in trees:
for feats, dependency_feat, relation in rst_tree.generate_relation_samples(
self.brown_clusters, level):
if level == 0:
yield vectorize(
feats, self.relation_feat_template_level_0
), self.relation_map[relation], dependency_feat[0]
if level == 1:
yield vectorize(
feats, self.relation_feat_template_level_1
), self.relation_map[relation], dependency_feat[0]
if level == 2:
yield vectorize(
feats, self.relation_feat_template_level_2
), self.relation_map[relation], dependency_feat[0]
def predict_probs(self, features, dependency_feat):
""" predict labels and rank the decision label with their confidence
value, output labels and probabilities
"""
vec = vectorize(features, self.feature_template)
# print(vec.toarray().shape)
# print(np.concatenate([vec.toarray(),dependency_feat],axis=1).shape)
feats = np.concatenate([vec.toarray(), dependency_feat], axis=1)
vals = self.classifier.decision_function(feats)
action_vals = {}
for idx in range(len(self.idxaction_map)):
action_vals[self.idxaction_map[idx]] = vals[0, idx]
sorted_actions = sorted(action_vals.items(), key=itemgetter(1), reverse=True)
return sorted_actions
def predict_probs(self, features):
""" predict labels and rank the decision label with their confidence
value, output labels and probabilities
"""
vec = vectorize(features, self.feature_template)
# vec = self.classifier.transform(vec)
try:
vals = self.classifier.decision_function(vec)
except:
# try:
vals = self.classifier.predict_proba(vec.toarray())
# except:
# raise NotImplementedError('Undefined classifier')
action_vals = {}
for idx in range(len(self.idxaction_map)):
action_vals[self.idxaction_map[idx]] = vals[0, idx]
sorted_actions = sorted(action_vals.items(),
key=itemgetter(1),
reverse=True)
return sorted_actions
def gen_action_train_data(self):
for rst_tree in self.rst_tree_instances:
for feats, action in rst_tree.generate_action_samples(
self.brown_clusters):
yield vectorize(
feats, self.action_feat_template), self.action_map[action]