def perceptron(self):
flag = True
if not self.iter_set:
self.iter += 1
for idx, sentence in enumerate(self.all_data, start=1):
full_graph = build_full_graph(len(sentence))
get_score_func = GetScore(idx, self.w, sentence,
self.feature_extractor)
digraph = Digraph(full_graph, get_score_func)
graph = build_real_graph(sentence)
mst = digraph.mst().successors
add_graph = {}
rm_graph = {}
for k in graph.keys():
add_graph[k] = [v for v in graph[k] if v not in mst[k]]
rm_graph[k] = [v for v in mst[k] if v not in graph[k]]
if any(add_graph.values()):
flag = False
temp_w = plus(
self.w,
self.graph_feature_extractor(idx, add_graph, sentence))
self.w = minus(
temp_w,
self.graph_feature_extractor(idx, rm_graph, sentence))
progress_bar(idx, len(self.all_data), "sentences")
return flag
def test_create_full_graph(self):
a = {0: [1, 2, 3], 1: [2, 3], 2: [1, 3], 3: [1, 2]}
graph = Digraph(a)
new_graph = graph.mst()
suc = new_graph.successors
self.make_vaild(suc)
suc = {0: suc}
m = self.Model(suc)
perc = StructPerceptron(m, directory=directory)
self.assertRaises(Exception, GraphUtil.create_full_graph(suc))
def test_perceptron(self):
n = 10
a = {0: [1, 2, 3], 1: [2, 3], 2: [1, 3], 3: [1, 2]}
graph = Digraph(a)
new_graph = graph.mst()
suc = new_graph.successors
suc = self.make_vaild(suc)
suc = {0: suc}
m = self.Model(suc)
perc = StructPerceptron(m, directory=directory)
self.assertRaises(Exception, perc.perceptron(n))
print(perc.perceptron(n).A)
def test_identical_dependency_tree(self):
a = {0: [1, 2, 3], 1: [2, 3], 2: [1, 3], 3: [1, 2]}
graph = Digraph(a)
new_graph = graph.mst()
suc = new_graph.successors
self.make_vaild(suc)
suc = {0: suc}
m = self.Model(suc)
perc = StructPerceptron(m, directory=directory)
GraphUtil.create_full_graph(suc)
full_graph = perc.full_graph[0]
self.assertEqual(True, GraphUtil.identical_dependency_tree(full_graph, a))
def infer(self, sentence):
full_graph = build_full_graph(len(sentence))
get_score_func = GetScore(-1, self.w, sentence, self.feature_extractor)
digraph = Digraph(full_graph, get_score_func)
mst = digraph.mst().successors
mst = inverse_graph(mst)
Word = namedtuple('Word', 'counter token pos head')
return [
Word(counter=word.counter,
token=word.token,
pos=word.pos,
head=mst[word.counter]) for word in sentence
]
def get_score(h, m):
return curr_G_wieghts.get((h, m), 0)
t_init = time.time()
######### run perceptron #########
for n in range(NUM_OF_PERCEPTRON_STEPS):
print ('Starting Perceptron Step - ' + str(n + 1))
for i in range(len(sent_word_list)):
curr_G = sent_graph_list[i]
# creates wieghts to each eadge
curr_G_wieghts = turn_edge_feats_to_wights(
edge_dict=sent_graph_edges_feats[i],
wights_vec=basic_feature_weights_vec)
graph = Digraph(curr_G, get_score)
if True == TRAIN_WITH_MST:
mst = graph.mst()
else:
mst = graph.greedy()
pred_heads = convert_chi_lui_output_to_list_of_heads(
sent_len=len(sent_word_list[i]),
mst=mst)
if len(pred_heads) != len(sent_head_list[i]):
print ('Perceptron prediction problem at sent ' + str(i) + ' pred len ' + str(pred_heads) + ' real heads len' + str(sent_head_list[i]))
if tuple(pred_heads) != tuple(sent_head_list[i]):
# perceptron correction step
pred_feat_idxs = get_all_feature_idxes_for_sent_and_head(
featurs_obj=featurs_basic_obj,
sent_words=sent_word_list[i],
sent_pos=sent_pos_list[i],
sent_heads=pred_heads)
# indexs of real X + Y features