def run_classify(self, train_path, test_path, train_set, test_set, output_path) :
loader = PickleMarket()
# read train
feature_names = loader.load_market(train_path)[0]
train_articles = list()
for type in train_set.split('#') :
path = train_path.replace(u'car', type)
train_articles.extend(loader.load_market(path)[1:])
train_dataset = np.array([np.array(article[1:30000], dtype=float) for article in train_articles])
print train_dataset.shape
train_label = np.array([np.array(int(article[-1])) for article in train_articles])
# read test
test_articles = list()
for type in test_set.split('#') :
path = test_path.replace(u'car', type)
test_articles.extend(loader.load_market(path)[1:])
test_dataset = np.array([np.array(article[1:30000]) for article in test_articles])
print test_dataset.shape
test_label = np.array([np.array(int(article[-1])) for article in test_articles])
# train cls
classifier = LrClassifier()
train_dataset = classifier.normalize(train_dataset, method='mapminmax')
test_dataset = classifier.normalize(test_dataset, method='mapminmax')
classifier.training(train_dataset, train_label, c=10, kernel='linear')
# test cls
test_prob = classifier.testing(test_dataset, type='prob')
test_class = classifier.testing(test_dataset, type='label')
evls, fprs, tprs = classifier.evaluation(test_label, test_prob, test_class)
print 'performance is', evls
ftprs = [[fpr, tprs[idx]] for idx, fpr in enumerate(fprs)]
file_operator = TextFileOperator()
file_operator.writing(ftprs, output_path)
print 'finish'
def run_convert_sentences(self, sentences_path, sentences_market_path) :
file_operator = TextFileOperator()
sentences = file_operator.reading(sentences_path)
sentences = [[word.split('<:>')[0] for word in sentence] for sentence in sentences]
loader = PickleMarket()
loader.dump_market(sentences, sentences_market_path)
print 'converting sentences finished ...'
def run_feature_select(self, article_market_path, dictionary_path, \
feature_market_path) :
loader = PickleMarket()
articles = loader.load_market(article_market_path)
[word2id, id2word] = loader.load_market(dictionary_path)
dim = len(word2id)
featuresets = list()
length = len(articles) - 1
for idx, article in enumerate(articles) :
feature = [0] * dim
for word in article['participle_title'] :
word = word.to_string()
if word in word2id :
feature[word2id[word]] += 1
for word in article['participle_content'] :
word = word.to_string()
if word in word2id :
feature[word2id[word]] += 1
featuresets.append([article['id']] + feature + [article['label']])
if idx % 100 == 0 :
print 'finish rate is %.2f%%\r' % (100.0*idx/length),
print 'finish rate is %.2f%%\r' % (100.0*idx/length)
file_operator = TextFileOperator()
# file_operator.writing(featuresets, feature_path)
loader.dump_market(featuresets, feature_market_path)
print 'finish'
def run_create_dictionary(self, article_market_path, dictionary_path, dict_set) :
loader = PickleMarket()
word2id, id2word = dict(), dict()
index = 0
for type in dict_set.split('#') :
path = article_market_path.replace(u'car', type)
articles = loader.load_market(path)
length = len(articles) - 1
for idx, article in enumerate(articles) :
for word in article['participle_title'] :
word = word.to_string()
if word not in word2id :
word2id[word] = index
id2word[index] = word
index += 1
for word in article['participle_content'] :
word = word.to_string()
if word not in word2id :
word2id[word] = index
id2word[index] = word
index += 1
if idx % 100 == 0 :
print 'finish rate is %.2f%%\r' % (100.0*idx/length),
print 'finish rate is %.2f%%\r' % (100.0*idx/length)
loader = PickleMarket()
loader.dump_market([word2id, id2word], dictionary_path)
def run_classify(self, train_path, test_path, train_set, test_set):
loader = PickleMarket()
# read train
feature_names = loader.load_market(train_path)[0]
train_articles = list()
for type in train_set.split("#"):
path = train_path.replace(u"car", type)
train_articles.extend(loader.load_market(path)[1:])
train_dataset = np.array([np.array(article[1:-1], dtype=float) for article in train_articles])
print train_dataset.shape
train_label = np.array([np.array(int(article[-1])) for article in train_articles])
# read test
test_articles = list()
for type in test_set.split("#"):
path = test_path.replace(u"car", type)
test_articles.extend(loader.load_market(path)[1:])
test_dataset = np.array([np.array(article[1:-1]) for article in test_articles])
print test_dataset.shape
test_label = np.array([np.array(int(article[-1])) for article in test_articles])
# train cls
classifier = SvmClassifier()
train_dataset = classifier.normalize(train_dataset, method="mapminmax")
test_dataset = classifier.normalize(test_dataset, method="mapminmax")
classifier.training(train_dataset, train_label, c=10, kernel="linear")
# test cls
test_prob = classifier.testing(test_dataset, type="prob")
test_class = classifier.testing(test_dataset, type="label")
print "performance is", classifier.evaluation(test_label, test_prob, test_class)
print "finish"
def run_feature_select(
self, article_market_path, pos_path, punc_path, klword_path, feature_path, feature_market_path
):
loader = PickleMarket()
pos_selector = selector.PosExtractor(pos_path, w=15, combined=False)
token_selector = selector.TokenExtractor(punc_path)
word_selector = selector.WordExtractor(klword_path, weight=1)
articles = loader.load_market(article_market_path)
length = len(articles) - 1
for idx, article in enumerate(articles):
article["features"] = list()
article["features"].extend(pos_selector.extract_feature_windows(article["participle_content"]))
article["features"].extend(
token_selector.extract_feature(
article["title"], article["content"], article["participle_title"], article["participle_content"]
)
)
article["features"].extend(
word_selector.extract_feature(article["participle_title"], article["participle_content"])
)
print "finish rate is %.2f%%\r" % (100.0 * idx / length),
print "finish rate is %.2f%%\r" % (100.0 * idx / length)
featuresets = [pos_selector.names + token_selector.names + word_selector.names]
featuresets.extend([[article["id"]] + article["features"] + [article["label"]] for article in articles])
file_operator = TextFileOperator()
file_operator.writing(featuresets, feature_path)
loader.dump_market(featuresets, feature_market_path)
print "finish"
def run_robot(self, tag_tree_path, sentences_market_path, tags_path) :
robot = Robot()
loader = PickleMarket()
file_operator = TextFileOperator()
cmd_list = file_operator.reading(tag_tree_path)
tag_tree = TagTree(cmd_list)
sentences = loader.load_market(sentences_market_path)
tags = loader.load_market(tags_path)
print 'start'
string = raw_input().decode('gb18030')
# string = u'我想要毛衣'
sentences = robot.question_and_answer(string, sentences, tags, tag_tree)
def run_create_word2vec(self, sentences_path, word_embedding_path, word_embedding_market_path) :
loader = PickleMarket()
sentences = list()
for type in [u'_car']:#, u'_finance', u'_web'] :
sentences.extend(loader.load_market(sentences_path + type))
print 'import finish ...'
embeddor = WordEmbed()
print sentences[0]
model = embeddor.word_to_vector(type='create', sentences=sentences[0:100], path=word_embedding_market_path)
data_list = embeddor.get_word2vec_model(model)
file_operator = TextFileOperator()
file_operator.writing(data_list, word_embedding_path)
print 'create word2vec finished ...'
def run_test(self, tag_tree_path, sentences_market_path, tags_path, \
tags_martket_path, untag_sentence_path) :
file_operator = TextFileOperator()
loader = PickleMarket()
sentences = loader.load_market(sentences_market_path)
cmd_list = file_operator.reading(tag_tree_path)
tag_tree = TagTree(cmd_list)
robot = Robot()
tags, tags_show, untag_sentences = robot.tag_sentences(tag_tree, sentences[0:])
loader = JsonMarket()
self.write_tags(sentences, tags_show, tags_path)
loader.dump_market(tags, tags_martket_path)
file_operator.writing(untag_sentences, untag_sentence_path)
# loader.dump_market(untag_sentences, sentences_market_path)
# print '%.2f%% article >= 2 tags' % (100.0 * len([tag for tag in tags_show if len(tag) >= 1]) / len(sentences))
print '%.2f%% article >= 3 tags' % (100.0 * len([tag for tag in tags_show if len(tag) >= 1]) / len(sentences))
def run_classify(self, train_path, test_path) :
loader = PickleMarket()
articles = list()
# for type in [u'car', u'finance', u'web'] :
# path = train_path.replace(u'all', type)
articles.extend(loader.load_market(train_path))
train_dataset = np.array([np.array(article[1:-1]) for article in articles])
print train_dataset.shape
train_label = np.array([np.array(int(article[-1])) for article in articles])
articles = loader.load_market(test_path)
test_dataset = np.array([np.array(article[1:-1]) for article in articles])
test_label = np.array([np.array(int(article[-1])) for article in articles])
classifier = SvmClassifier()
train_dataset = classifier.normalize(train_dataset, method='mapminmax')
test_dataset = classifier.normalize(test_dataset, method='mapminmax')
classifier.training(train_dataset, train_label, cset=range(10, 100, 10), kernel='linear')
test_prob = classifier.testing(test_dataset, type='prob')
test_class = classifier.testing(test_dataset, type='label')
print 'performance is', classifier.evaluation(test_label, test_prob, test_class)
print 'finish'
def run_feature_select(self, article_market_path, pos_path, punc_path, klword_path, \
feature_path, feature_market_path) :
loader = PickleMarket()
pos_selector = selector.PosExtractor(pos_path, w=5, combined=True)
token_selector = selector.TokenExtractor(punc_path)
word_selector = selector.WordExtractor(klword_path, weight=1)
articles = loader.load_market(article_market_path)
length = len(articles) - 1
for idx, article in enumerate(articles) :
article['features'] = list()
article['features'].extend(pos_selector.extract_feature(article['participle_content']))
article['features'].extend(token_selector.extract_feature(article['title'], article['content'], \
article['participle_title'], \
article['participle_content']))
article['features'].extend(word_selector.extract_feature(article['participle_title'], \
article['participle_content']))
print 'finish rate is %.2f%%\r' % (100.0*idx/length),
print 'finish rate is %.2f%%\r' % (100.0*idx/length)
featuresets = [[article['id']] + article['features'] + [article['label']] for article in articles]
file_operator = TextFileOperator()
file_operator.writing(featuresets, feature_path)
loader.dump_market(featuresets, feature_market_path)
print 'finish'
def run_tag_sentences(self, tag_tree_path, sentences_market_path, tags_martket_path, dict_market_path) :
file_operator = TextFileOperator()
loader = PickleMarket()
sentences = loader.load_market(sentences_market_path)
cmd_list = file_operator.reading(tag_tree_path)
tag_tree = TagTree(cmd_list)
robot = Robot()
tags, tags_show, untag_sentences = robot.tag_sentences(tag_tree, sentences[0:])
loader = JsonMarket()
loader.dump_market(tags, tags_martket_path)
loader.dump_market(tag_tree.dict_tuple, dict_market_path)
print '%.2f%% article >= 1 tags, number is, %d.' \
% (100.0 * len([tag for tag in tags_show if len(tag) >= 1]) / len(sentences)) \
% len([tag for tag in tags_show if len(tag) >= 1])
def run_convert_sentences(self, sentences_path, sentences_market_path) :
file_operator = TextFileOperator()
sentences = self.read_sentences(sentences_path)
loader = PickleMarket()
loader.dump_market(sentences, sentences_market_path)
def run_optimize_params(self, train_article_market_path, test_article_market_path, \
pos_path, punc_path, klword_path, logger_path) :
loader = PickleMarket()
logger = list()
logger.append(['w', 'combined', 'weight', 'kernel', 'c', 'norm', 'car_car', \
'car_finance', 'car_web', 'finance_car', 'finance_finance', 'finance_web', \
'web_car', 'web_fiannce', 'web_web', 'merge_car', 'merge_finance', 'merge_web'])
domains = [u'car', u'finance', u'web']
wset = [5, 10, 15, 20]
combinedset = [True, False]
weightset = [1, 2, 5]
kernelset = ['linear', 'poly', 'rbf']
cset = [range(10, 100, 10), range(100, 1000, 100)]
normset = ['mapminmax', 'zscore']
token_selector = selector.TokenExtractor(punc_path)
for w in wset :
for combined in combinedset :
pos_selector = selector.PosExtractor(pos_path, w=w, combined=combined)
for weight in weightset :
word_selector = selector.WordExtractor(klword_path, weight=weight)
train_featuresets, test_featuresets = list(), list()
for step in range(len(domains)) :
train_featuresets.append(list())
test_featuresets.append(list())
for index, domain in enumerate(domains) :
train_articles = loader.load_market(train_article_market_path.replace(u'all', domain))
test_articles = loader.load_market(test_article_market_path.replace(u'all', domain))
# train
length = len(train_articles) - 1
for idx, article in enumerate(train_articles) :
article['features'] = list()
article['features'].extend(pos_selector.extract_feature(article['participle_content']))
article['features'].extend(token_selector.extract_feature(article['title'], article['content'], \
article['participle_title'], \
article['participle_content']))
article['features'].extend(word_selector.extract_feature(article['participle_title'], \
article['participle_content']))
print 'finish rate is %.2f%%\r' % (100.0*idx/length),
print 'finish rate is %.2f%%\r' % (100.0*idx/length)
train_featuresets[index] = [[article['id']] + article['features'] + [article['label']] for article in train_articles]
# test
length = len(test_articles) - 1
for idx, article in enumerate(test_articles) :
article['features'] = list()
article['features'].extend(pos_selector.extract_feature(article['participle_content']))
article['features'].extend(token_selector.extract_feature(article['title'], article['content'], \
article['participle_title'], \
article['participle_content']))
article['features'].extend(word_selector.extract_feature(article['participle_title'], \
article['participle_content']))
print 'finish rate is %.2f%%\r' % (100.0*idx/length),
print 'finish rate is %.2f%%\r' % (100.0*idx/length)
test_featuresets[index] = [[article['id']] + article['features'] + [article['label']] for article in test_articles]
for kernel in kernelset :
for c in cset :
for norm in normset :
evl = list()
for train_idx in range(0, len(domains)) :
for test_idx in range(0, len(domains)) :
train_dataset = np.array([np.array(article[1:-1]) for article in train_featuresets[train_idx]])
train_label = np.array([np.array(int(article[-1])) for article in train_featuresets[train_idx]])
test_dataset = np.array([np.array(article[1:-1]) for article in test_featuresets[test_idx]])
test_label = np.array([np.array(int(article[-1])) for article in test_featuresets[test_idx]])
classifier = SvmClassifier()
train_dataset = classifier.normalize(train_dataset, method=norm)
test_dataset = classifier.normalize(test_dataset, method=norm)
classifier.training(train_dataset, train_label, cset=c, kernel=kernel)
test_prob = classifier.testing(test_dataset, type='prob')
test_class = classifier.testing(test_dataset, type='label')
evl.append(classifier.evaluation(test_label, test_prob, test_class)[1])
print 'single finished ...'
# merge
articles = list()
for train_idx in range(0, len(domains)) :
articles.extend(train_featuresets[train_idx])
train_dataset = np.array([np.array(article[1:-1]) for article in articles])
train_label = np.array([np.array(int(article[-1])) for article in articles])
for test_idx in range(0, len(domains)) :
test_dataset = np.array([np.array(article[1:-1]) for article in test_featuresets[test_idx]])
test_label = np.array([np.array(int(article[-1])) for article in test_featuresets[test_idx]])
classifier = SvmClassifier()
train_dataset = classifier.normalize(train_dataset, method=norm)
test_dataset = classifier.normalize(test_dataset, method=norm)
classifier.training(train_dataset, train_label, cset=c, kernel=kernel)
test_prob = classifier.testing(test_dataset, type='prob')
test_class = classifier.testing(test_dataset, type='label')
evl.append(classifier.evaluation(test_label, test_prob, test_class)[1])
print 'merge finished ...'
print 'performance is', 1.0*sum(evl)/len(evl)
log = [w, combined, weight, kernel, c[0], norm]
log.extend(evl)
logger.append(log)
file_operator = TextFileOperator()
file_operator.writing(logger, logger_path)
print 'finish'
def run_convert_article(self, article_path, article_market_path) :
articles = self.read_article(article_path)
# articles = self.read_participle(articles, participle_path)
loader = PickleMarket()
loader.dump_market(articles, article_market_path)
print 'finish.'
def run_optimize_params(
self, train_article_market_path, test_article_market_path, pos_path, punc_path, klword_path, logger_path
):
loader = PickleMarket()
logger = list()
logger.append(
[
"w",
"combined",
"weight",
"kernel",
"c",
"norm",
"car_car",
"car_finance",
"car_web",
"finance_car",
"finance_finance",
"finance_web",
"web_car",
"web_fiannce",
"web_web",
"merge_car",
"merge_finance",
"merge_web",
]
)
domains = [u"car", u"finance", u"web"]
wset = [5, 10, 15, 20]
combinedset = [True, False]
weightset = [1, 2, 5]
kernelset = ["linear", "poly", "rbf"]
cset = [range(10, 100, 10), range(100, 1000, 100)]
normset = ["mapminmax", "zscore"]
token_selector = selector.TokenExtractor(punc_path)
for w in wset:
for combined in combinedset:
pos_selector = selector.PosExtractor(pos_path, w=w, combined=combined)
for weight in weightset:
word_selector = selector.WordExtractor(klword_path, weight=weight)
train_featuresets, test_featuresets = list(), list()
for step in range(len(domains)):
train_featuresets.append(list())
test_featuresets.append(list())
for index, domain in enumerate(domains):
train_articles = loader.load_market(train_article_market_path.replace(u"all", domain))
test_articles = loader.load_market(test_article_market_path.replace(u"all", domain))
# train
length = len(train_articles) - 1
for idx, article in enumerate(train_articles):
article["features"] = list()
article["features"].extend(pos_selector.extract_feature(article["participle_content"]))
article["features"].extend(
token_selector.extract_feature(
article["title"],
article["content"],
article["participle_title"],
article["participle_content"],
)
)
article["features"].extend(
word_selector.extract_feature(
article["participle_title"], article["participle_content"]
)
)
print "finish rate is %.2f%%\r" % (100.0 * idx / length),
print "finish rate is %.2f%%\r" % (100.0 * idx / length)
train_featuresets[index] = [
[article["id"]] + article["features"] + [article["label"]] for article in train_articles
]
# test
length = len(test_articles) - 1
for idx, article in enumerate(test_articles):
article["features"] = list()
article["features"].extend(pos_selector.extract_feature(article["participle_content"]))
article["features"].extend(
token_selector.extract_feature(
article["title"],
article["content"],
article["participle_title"],
article["participle_content"],
)
)
article["features"].extend(
word_selector.extract_feature(
article["participle_title"], article["participle_content"]
)
)
print "finish rate is %.2f%%\r" % (100.0 * idx / length),
print "finish rate is %.2f%%\r" % (100.0 * idx / length)
test_featuresets[index] = [
[article["id"]] + article["features"] + [article["label"]] for article in test_articles
]
for kernel in kernelset:
for c in cset:
for norm in normset:
evl = list()
for train_idx in range(0, len(domains)):
for test_idx in range(0, len(domains)):
train_dataset = np.array(
[np.array(article[1:-1]) for article in train_featuresets[train_idx]]
)
train_label = np.array(
[np.array(int(article[-1])) for article in train_featuresets[train_idx]]
)
test_dataset = np.array(
[np.array(article[1:-1]) for article in test_featuresets[test_idx]]
)
test_label = np.array(
[np.array(int(article[-1])) for article in test_featuresets[test_idx]]
)
classifier = SvmClassifier()
train_dataset = classifier.normalize(train_dataset, method=norm)
test_dataset = classifier.normalize(test_dataset, method=norm)
classifier.training(train_dataset, train_label, cset=c, kernel=kernel)
test_prob = classifier.testing(test_dataset, type="prob")
test_class = classifier.testing(test_dataset, type="label")
evl.append(classifier.evaluation(test_label, test_prob, test_class)[1])
print "single finished ..."
# merge
articles = list()
for train_idx in range(0, len(domains)):
articles.extend(train_featuresets[train_idx])
train_dataset = np.array([np.array(article[1:-1]) for article in articles])
train_label = np.array([np.array(int(article[-1])) for article in articles])
for test_idx in range(0, len(domains)):
test_dataset = np.array(
[np.array(article[1:-1]) for article in test_featuresets[test_idx]]
)
test_label = np.array(
[np.array(int(article[-1])) for article in test_featuresets[test_idx]]
)
classifier = SvmClassifier()
train_dataset = classifier.normalize(train_dataset, method=norm)
test_dataset = classifier.normalize(test_dataset, method=norm)
classifier.training(train_dataset, train_label, cset=c, kernel=kernel)
test_prob = classifier.testing(test_dataset, type="prob")
test_class = classifier.testing(test_dataset, type="label")
evl.append(classifier.evaluation(test_label, test_prob, test_class)[1])
print "merge finished ..."
print "performance is", 1.0 * sum(evl) / len(evl)
log = [w, combined, weight, kernel, c[0], norm]
log.extend(evl)
logger.append(log)
file_operator = TextFileOperator()
file_operator.writing(logger, logger_path)
print "finish"
