def real_test(self):
self._test_xs, self._test_ys = ST.load_data(self.test_path)
ST.replace_url(self._test_xs, fill='H')
ST.replace_target(self._test_xs, fill='T')
#x_y = [(self.discret_txt(txt), y) for txt, y in zip(self._test_xs, self._test_ys)]
test_mat = self.build_sparse_X(self._test_xs)
self.accuracy(test_mat, self._test_ys)
def parse_stats_emoticon(in_path='../stats/train_public_stats',
out_path='../stats/emoticon_debug'):
'''
Study emoticon info from public states
'''
st = time.time()
icon2cnt = {}
config = {"row_num": None}
lines = ST.load_raw_data(in_path, **config)
icon_line_cnt = 0
for txt in lines:
icons = ST._retrieve_emoticon(txt)
if not icons:
continue
icon_line_cnt += 1
for icon in icons:
icon2cnt.setdefault(icon, 0)
icon2cnt[icon] += 1
if os.path.exists(out_path):
os.system('rm %s' % out_path)
icons = icon2cnt.keys()
icons = sorted(icons, key=lambda x: icon2cnt[x], reverse=True)
for icon in icons:
cnt = icon2cnt[icon]
write(out_path, 'a', '%s:%s\n' % (icon, cnt))
linfo('end parse emoticons. total lines: %s.icon lines: %s. icons:%s.' %
(len(lines), icon_line_cnt, len(icons)))
def parse_stats_emoticon(in_path='../stats/train_public_stats', out_path='../stats/emoticon_debug'):
'''
Study emoticon info from public states
'''
st = time.time()
icon2cnt = {}
config={"row_num":None}
lines = ST.load_raw_data(in_path, **config)
icon_line_cnt = 0
for txt in lines:
icons = ST._retrieve_emoticon(txt)
if not icons:
continue
icon_line_cnt += 1
for icon in icons:
icon2cnt.setdefault(icon, 0)
icon2cnt[icon] += 1
if os.path.exists(out_path):
os.system('rm %s' % out_path)
icons = icon2cnt.keys()
icons = sorted(icons, key=lambda x: icon2cnt[x], reverse=True)
for icon in icons:
cnt = icon2cnt[icon]
write(out_path, 'a', '%s:%s\n' % (icon, cnt))
linfo('end parse emoticons. total lines: %s.icon lines: %s. icons:%s.' % (len(lines), icon_line_cnt, len(icons)))
def debug(self):
ws_1 = set(self.batch_extract_feature().keys())
ST.remove_emoticon(self._train_xs)
ws_2 = set(self.batch_extract_feature().keys())
linfo('uni_bi_icon_feature_cnt: %s. no_icon: %s' % (len(ws_1), len(ws_2)))
rms = ws_2 - ws_1
for x in rms:
print x
def debug(self):
ws_1 = set(self.batch_extract_feature().keys())
ST.remove_emoticon(self._train_xs)
ws_2 = set(self.batch_extract_feature().keys())
linfo('uni_bi_icon_feature_cnt: %s. no_icon: %s' %
(len(ws_1), len(ws_2)))
rms = ws_2 - ws_1
for x in rms:
print x
def train_discret_model(self, **config):
linfo('begin train helper discret model: %s' % config)
if not config['emoticon']:
ST.remove_emoticon(self._train_xs)
if not config['parenthesis']:
ST.remove_parenthesis(self._train_xs)
self.txt2bags = {}
self.w2id = self.batch_extract_feature()
linfo('end train helper discret model')
def train_discret_model(self, **config):
linfo('begin train helper discret model: %s' % config)
if not config['emoticon']:
ST.remove_emoticon(self._train_xs)
if not config['parenthesis']:
ST.remove_parenthesis(self._train_xs)
self.txt2bags = {}
self.w2id = self.batch_extract_feature()
linfo('end train helper discret model')
def format_test(self, emoticon=True, parenthesis=True):
test_path='../test_data/%s_test_data' % self.classifier_type
self._test_xs, self._test_ys = ST.load_data(test_path)
linfo('begin preprocess test data, then sparse')
self._raw_test_xs, self._test_xs = ST.preprocess(self._test_xs)
#ST.replace_url(self._test_xs, fill='H')
#ST.replace_target(self._test_xs, fill='T')
self._test_ys = map(lambda x:self.tag2index[x], self._test_ys)
self.format_sparse(self._test_xs, self._test_ys, '%s/test_data/%s%s_sparse_test_data_%s' % (project_dir, self.flag_prefix, self.classifier_type, 'icon' if emoticon else 'no_icon'))
def train(self):
self._train_xs, self._train_ys = ST.load_data(self._path)
if not self._emoticon:
ST.remove_emoticon(self._train_xs)
self.gram2gid = self._discretize_gram2gid()
X = self.build_sparse_X(self._train_xs)
self.clf.fit(X, self._train_ys)
self.real_test()
def parse_emoticon_stats(in_path='../stats/train_public_stats',
out_path='../stats/train_data_dg'):
'''
Parse states with selected emocicons.
Dumps or visualise
'''
st = time.time()
pos_icons, neg_icons = load_emoticon()
icon2stat = {}
lines = ST.load_raw_data(in_path)
for txt in lines:
if any([x in txt for x in excludes]):
continue
icons = ST._retrieve_emoticon(txt)
if not icons:
continue
dis_match = filter(lambda x: x not in pos_icons and x not in neg_icons,
icons)
if dis_match:
if len(set(dis_match)) >= 2:
continue
pos_match = filter(lambda x: x in txt, pos_icons)
neg_match = filter(lambda x: x in txt, neg_icons)
if (pos_match and neg_match) or (not pos_match and not neg_match):
continue
if pos_match:
for icon in pos_match:
icon2stat.setdefault(icon, [])
icon2stat[icon].append(txt)
break
if neg_match:
for icon in neg_match:
icon2stat.setdefault(icon, [])
icon2stat[icon].append(txt)
break
write = ET.write_file
if os.path.exists(out_path):
os.system('rm %s' % out_path)
pos_cnt = sum([len(icon2stat.get(x, [])) for x in pos_icons])
neg_cnt = sum([len(icon2stat.get(x, [])) for x in neg_icons])
icons = copy.copy(pos_icons)
icons.extend(neg_icons)
write(
out_path, 'a',
'----------------\ntotal_cnt: %s. pos_cnt: %s. neg_cnt: %s. time used: %.2fs\n'
% (len(lines), pos_cnt, neg_cnt, time.time() - st))
for icon in icons:
stats = icon2stat.get(icon, [])
#write(out_path, 'a', '--------------------------------------\nicon: %s. stats_cnt: %s\n' % (icon, len(stats)))
for stat in stats:
dic = {'%s' % ('P' if icon in pos_icons else 'N'): stat}
write(out_path, 'a', '%s\n' % json.dumps(dic))
def train(self, icon=True, cross=False):
#word2cnt = BayesClassifier.Word2Cnt()
#txt = '今天天气就是棒[哈哈] [太阳] [飞起来]#'
#return
#self._load_data()
#self._replace_url(fill=True)
self._train_xs, self._train_ys = ST.load_data(self._train_path)
ST.replace_url(self._train_xs, fill=True)
if not icon:
ST.remove_emoticon(self._train_xs)
self._train(cross_validation=cross)
def train(self, icon=True, cross=False):
#word2cnt = BayesClassifier.Word2Cnt()
#txt = '今天天气就是棒[哈哈] [太阳] [飞起来]#'
#return
#self._load_data()
#self._replace_url(fill=True)
self._train_xs, self._train_ys = ST.load_data(self._train_path)
ST.replace_url(self._train_xs, fill=True)
if not icon:
ST.remove_emoticon(self._train_xs)
self._train(cross_validation=cross)
def format_test(self, emoticon=True, parenthesis=True):
test_path = '../test_data/%s_test_data' % self.classifier_type
self._test_xs, self._test_ys = ST.load_data(test_path)
linfo('begin preprocess test data, then sparse')
self._raw_test_xs, self._test_xs = ST.preprocess(self._test_xs)
#ST.replace_url(self._test_xs, fill='H')
#ST.replace_target(self._test_xs, fill='T')
self._test_ys = map(lambda x: self.tag2index[x], self._test_ys)
self.format_sparse(
self._test_xs, self._test_ys,
'%s/test_data/%s%s_sparse_test_data_%s' %
(project_dir, self.flag_prefix, self.classifier_type,
'icon' if emoticon else 'no_icon'))
def parse_emoticon_stats(in_path='../stats/train_public_stats', out_path='../stats/train_data_dg'):
'''
Parse states with selected emocicons.
Dumps or visualise
'''
st = time.time()
pos_icons, neg_icons = load_emoticon()
icon2stat= {}
lines = ST.load_raw_data(in_path)
for txt in lines:
if any([x in txt for x in excludes]):
continue
icons = ST._retrieve_emoticon(txt)
if not icons:
continue
dis_match = filter(lambda x:x not in pos_icons and x not in neg_icons, icons)
if dis_match:
if len(set(dis_match)) >= 2:
continue
pos_match = filter(lambda x: x in txt, pos_icons)
neg_match = filter(lambda x: x in txt, neg_icons)
if (pos_match and neg_match) or (not pos_match and not neg_match):
continue
if pos_match:
for icon in pos_match:
icon2stat.setdefault(icon, [])
icon2stat[icon].append(txt)
break
if neg_match:
for icon in neg_match:
icon2stat.setdefault(icon, [])
icon2stat[icon].append(txt)
break
write = ET.write_file
if os.path.exists(out_path):
os.system('rm %s' % out_path)
pos_cnt = sum([len(icon2stat.get(x, [])) for x in pos_icons])
neg_cnt = sum([len(icon2stat.get(x, [])) for x in neg_icons])
icons = copy.copy(pos_icons)
icons.extend(neg_icons)
write(out_path, 'a', '----------------\ntotal_cnt: %s. pos_cnt: %s. neg_cnt: %s. time used: %.2fs\n' % (len(lines), pos_cnt, neg_cnt, time.time()-st))
for icon in icons:
stats = icon2stat.get(icon, [])
#write(out_path, 'a', '--------------------------------------\nicon: %s. stats_cnt: %s\n' % (icon, len(stats)))
for stat in stats:
dic = {'%s' % ('P' if icon in pos_icons else 'N'):stat }
write(out_path, 'a', '%s\n' % json.dumps(dic))
def test():
obj_stats_path = '../train_data/stat_obj_train_data'
out_path = '../train_data/Dg_obj_stats'
txts = []
with open(obj_stats_path, 'r') as f:
for line in f:
dic = json.loads(line.strip())
tag, txt = dic.items()[0]
txts.append(txt)
linfo('obj stats count: %s' % (len(txts)))
ST.replace_url(txts, fill='H')
ST.replace_target(txts, fill='T')
for x in txts:
dic = {'O':x}
write(out_path, 'a', '%s\n' % json.dumps(dic))
def test():
obj_stats_path = '../train_data/stat_obj_train_data'
out_path = '../train_data/Dg_obj_stats'
txts = []
with open(obj_stats_path, 'r') as f:
for line in f:
dic = json.loads(line.strip())
tag, txt = dic.items()[0]
txts.append(txt)
linfo('obj stats count: %s' % (len(txts)))
ST.replace_url(txts, fill='H')
ST.replace_target(txts, fill='T')
for x in txts:
dic = {'O': x}
write(out_path, 'a', '%s\n' % json.dumps(dic))
def parse_topic_public_stats(in_path='../stats/train_public_stats',
out_path='../test_data/topic_test_data'):
st_t = time.time()
topic_cnt, total_cnt = 0, 0
topic2txt = {}
with open(in_path, 'r') as f:
for line in f:
total_cnt += 1
dic = json.loads(line.strip())
txt = dic['text']
topic = ST.parse_topic(txt)
if not topic:
continue
topic2txt.setdefault(topic, list())
topic2txt[topic].append(txt)
topics = sorted(topic2txt.keys(),
key=lambda x: len(topic2txt[x]),
reverse=True)
for t in topics:
txts = topic2txt[t]
if len(txts) > 7000:
continue
#print t, topic2txt[t]
if len(txts) < 200:
break
for txt in txts:
dic = {t: txt}
ET.write_file(out_path, 'a', '%s\n' % json.dumps(dic))
print 'total cnt: %s. topic stats cnt: %s' % (total_cnt, topic_cnt)
print 'topic cnt: %s' % len(topic2txt)
print 'time used: %.2f' % (time.time() - st_t)
def update_profile_topic(self, raw_stats, tags):
for txt, tag in zip(raw_stats, tags):
topic = ST.parse_topic(txt)
if not topic:
continue
self.profile_topic.setdefault(topic, {"P":0,"N":0,"O":0})
self.profile_topic[topic][tag] += 1
def parse_topic_public_stats(in_path='../stats/train_public_stats',out_path='../test_data/topic_test_data'):
st_t = time.time()
topic_cnt, total_cnt = 0, 0
topic2txt = {}
with open(in_path, 'r') as f:
for line in f:
total_cnt += 1
dic = json.loads(line.strip())
txt = dic['text']
topic = ST.parse_topic(txt)
if not topic:
continue
topic2txt.setdefault(topic, list())
topic2txt[topic].append(txt)
topics = sorted(topic2txt.keys(), key=lambda x: len(topic2txt[x]), reverse=True)
for t in topics:
txts = topic2txt[t]
if len(txts) > 7000:
continue
#print t, topic2txt[t]
if len(txts) < 200:
break
for txt in txts:
dic = {t:txt}
ET.write_file(out_path, 'a', '%s\n' % json.dumps(dic))
print 'total cnt: %s. topic stats cnt: %s' % (total_cnt, topic_cnt)
print 'topic cnt: %s' % len(topic2txt)
print 'time used: %.2f' % (time.time() - st_t)
def ProfileRawData(path='../stats/train_public_stats'):
'''
calculate user, url, retweet, topic, redundant stat
'''
user_cnt, url_cnt, retweet_cnt, topic_cnt, redundant = (0, 0, 0, 0, 0)
st = time.time()
lines = ST.load_raw_data(path, replace_enabled=False, row_num=None)
w2c = {}
for txt in lines:
for x in txt:
w2c.setdefault(x, 0)
w2c[x] += 1
print 'word cnt', len(w2c)
out_path = 'word2cnt'
ET.write_file(out_path, 'w', '')
for w,c in w2c.items():
if w == ',':
print 'special word: %s. cnt %s' % (w, c)
continue
ET.write_file(out_path, 'a', '%s,%s\n' % (w, c))
return
for txt in lines:
if '@' in txt:
user_cnt += 1
if 'http' in txt:
url_cnt += 1
if '#' in txt:
st_i = txt.find('#')
if txt.find('#', st_i+1) != -1:
topic_cnt += 1
print 'user_cnt', user_cnt
print 'url_cnt', url_cnt
print 'topic_cnt', topic_cnt
print 'time used', time.time() - st
def online_run(self, interval=10, peroid=0.5, quiet=True):
'''
return value: [(city, stat)...]
'''
stats_set = set()
stats = []
now = peroid
cnt = 0
while now < interval:
try:
rsp = self.retrieve('on', quiet=quiet)
cnt += 1
if rsp:
for dic in rsp:
if dic['id'] not in stats_set:
city = ST.parse_spatial(dic)
item = (city, dic['text'])
stats.append(item)
stats_set.add(dic['id'])
except Exception as e:
logging.exception(e)
now += peroid
time.sleep(peroid*60)
linfo('online analysis %s new stats retrieved. retrieve cnt: %s' % (len(stats), cnt))
return stats
def _predict(self, txt, train_w2c, train_t2c, debug=False, emoticon=True):
#if emoticon and 'emoticon' not in self._ngrams_config:
# self._ngrams_config.append('emoticon')
#elif not emoticon and 'emoticon' in self._ngrams_config:
# self._ngrams_config = filter(lambda x: x != 'emoticon', self._ngrams_config)
#grams = self._retrieve_feature(txt)
grams = ST.retrieve_feature(txt,
feature_extract_config=self._ngrams_config,
gram_icon_mixed=emoticon)
if debug:
linfo('begin debug case: %s' % txt)
tag2score = {"P": 0, "N": 0, "O": 0}
for w in grams:
for tag in tag2score:
if not train_t2c[tag]:
continue
score = self._cal_likelihood(train_w2c[tag].get(w, 0),
train_t2c[tag])
tag2score[tag] += score
if debug:
linfo(
'DEBUG probability for gram %s when given tag %s is: %.4f. gram cnt: %s.tag cnt: %s'
% (w, tag, score, train_w2c[tag].get(
w, 0), train_t2c[tag]))
pred_tag = sorted(tag2score.keys(),
key=lambda x: tag2score[x],
reverse=True)[0]
if debug:
linfo('predict tag2score: %s' % tag2score)
return pred_tag
def run(self, detail=False):
self.train()
for name, stat in self.stats:
try:
for clf, path in self.classifiers:
linfo('----------roundly predict start-----------')
raw_stat, stat = ST.preprocess(stat)
union = [(raw,new) for raw, new in zip(raw_stat, stat) if new]
raw_stat = map(lambda x:x[0], union)
stat = map(lambda x:x[1], union)
pred_tags = clf.predict(stat)
if not pred_tags or len(pred_tags) != len(stat):
raise Exception('Predict Results Exception')
tag2dist = self.cal_tag_dist(pred_tags)
linfo('%s-roundly online sentiment distribution: %s' % (clf, tag2dist))
save(path, 'a', '%s\t%s\t%s\n' % (name, json.dumps(tag2dist), len(stat)))
if detail:
detail_path = '%s%s' % (stats_predict_detail_prefix, name)
if os.path.exists(detail_path):
os.system('rm %s' % detail_path)
for tag, txt in zip(pred_tags, raw_stat):
ET.write_file(detail_path, 'a', '%s -%s\n' % (tag, txt))
#print tag, '-%s' % txt
linfo('----------roundly predict end-----------')
except Exception as e:
lexcept('Unknown exception %s' % e)
def update_profile_topic(self, raw_stats, tags):
for txt, tag in zip(raw_stats, tags):
topic = ST.parse_topic(txt)
if not topic:
continue
self.profile_topic.setdefault(topic, {"P": 0, "N": 0, "O": 0})
self.profile_topic[topic][tag] += 1
def ProfileRawData(path='../stats/train_public_stats'):
'''
calculate user, url, retweet, topic, redundant stat
'''
user_cnt, url_cnt, retweet_cnt, topic_cnt, redundant = (0, 0, 0, 0, 0)
st = time.time()
lines = ST.load_raw_data(path, replace_enabled=False, row_num=None)
w2c = {}
for txt in lines:
for x in txt:
w2c.setdefault(x, 0)
w2c[x] += 1
print 'word cnt', len(w2c)
out_path = 'word2cnt'
ET.write_file(out_path, 'w', '')
for w, c in w2c.items():
if w == ',':
print 'special word: %s. cnt %s' % (w, c)
continue
ET.write_file(out_path, 'a', '%s,%s\n' % (w, c))
return
for txt in lines:
if '@' in txt:
user_cnt += 1
if 'http' in txt:
url_cnt += 1
if '#' in txt:
st_i = txt.find('#')
if txt.find('#', st_i + 1) != -1:
topic_cnt += 1
print 'user_cnt', user_cnt
print 'url_cnt', url_cnt
print 'topic_cnt', topic_cnt
print 'time used', time.time() - st
def parse_city_public_stats(in_path='../stats/train_public_stats', out_path='../test_data/city_test_data'):
st_t = time.time()
city2txt = {}
city_stat_cnt, total_cnt = 0, 0
stat_ids = set()
txts_upperbound = 1000
with open(in_path, 'r') as f:
for line in f:
total_cnt += 1
dic = json.loads(line.strip())
if dic['id'] in stat_ids:
continue
else:
stat_ids.add(dic['id'])
city = ST.parse_spatial(dic)
if not city:
continue
city2txt.setdefault(city, list())
if len(city2txt[city]) >= txts_upperbound:
continue
city2txt[city].append(dic['text'])
locs = sorted(city2txt.keys(), key=lambda x: len(city2txt[x]), reverse=True)
print 'city_stat_cnt', city_stat_cnt
print 'total_cnt', total_cnt
print 'time used: %.2f' % (time.time() - st_t)
citys = sorted(city2txt.keys())
#for x in citys:
# print x, len(city2txt[x])
if os.path.exists(out_path):
os.system('rm %s' % out_path)
for x in locs:
for txt in city2txt[x]:
dic={x:txt}
ET.write_file(out_path, 'a', '%s\n' % json.dumps(dic))
def run(self, sample_enabled=False, profile_enabled=False):
for csf, path in self.classifiers:
csf.train()
action_day = ET.format_time(time.localtime())[:10]
action_total_cnt = 0
if profile_enabled:
self.reset_profile()
while True:
try:
stats = self.psr.online_run(interval=10)
if not stats:
continue
linfo('-------roundly analysis----------')
citys = map(lambda x: x[0], stats)
stats = map(lambda x: x[1], stats)
raw_stats, stats = ST.preprocess(stats)
valid_ids = [i for i, txt in enumerate(stats) if txt]
stats = map(lambda i: stats[i], valid_ids)
raw_stats = map(lambda i: raw_stats[i], valid_ids)
citys = map(lambda i: citys[i], valid_ids)
f_t = ET.format_time(time.localtime())
if sample_enabled:
sample_path = '%srealtime_%s' % (
sample_prefix, f_t.replace(' ', '').replace(
'-', '').replace(':', ''))
ET.write_file(sample_path, 'a',
'%s\n' % json.dumps(raw_stats[:300]))
#only one model supported at the same time now.
for clf, path in self.classifiers:
tag2cnt = {'P': 0, 'N': 0, 'O': 0}
pred_tags = clf.predict(stats)
for tag in pred_tags:
tag2cnt[tag] += 1
tag2dist = {
tag: cnt * 1.0 / len(stats)
for tag, cnt in tag2cnt.items()
}
linfo('%s-roundly online sentiment distribution: %s' %
(clf, tag2dist))
f_time = ET.format_time(time.localtime())
today = f_time[:10]
action_total_cnt = (
action_total_cnt + len(pred_tags)
) if today == action_day else len(pred_tags)
save(
path, 'a', '%s\t%s\t%s\n' %
(f_time, json.dumps(tag2dist), len(stats)))
if profile_enabled:
self.update_profile_spatial(citys, pred_tags)
self.update_profile_topic(raw_stats, pred_tags)
if today != action_day:
self.save_profile(action_day)
self.reset_profile()
action_day = today
break
except Exception as e:
lexcept('Unknown exception %s' % e)
def _feature_encoding(self, txt):
bags = ST.retrieve_feature(txt, feature_extract_config=self._feature_extract_config)
#fs = {x:0 for x in gram2gid}
fs = {}
for gram in bags:
if gram in self.gram2gid:
fs[self.gram2gid[gram]] = 1
return fs
def get_feature(self, txt, cache=False):
if txt in self.txt2bags:
bags = self.txt2bags[txt]
else:
bags = ST.retrieve_feature(txt, feature_extract_config=self._feature_extract_config)
if cache:
self.txt2bags[txt] = bags
return bags
def _discretize_gram2gid(self):
w2id = {}
for txt in self._train_xs:
bags = ST.retrieve_feature(txt, feature_extract_config=self._feature_extract_config)
for w in bags:
if w not in w2id:
w2id[w] = len(w2id) + 1
linfo('grams cnt: %s' % len(w2id))
return w2id
def discret_txt(self, txt):
fs = [0 for x in range(len(self.gram2gid))]
bags = ST.retrieve_feature(
txt, feature_extract_config=self._feature_extract_config)
for w in bags:
if w in self.gram2gid:
wid = self.gram2gid[w]
fs[wid] = 1
return fs
def get_feature(self, txt, cache=False):
if txt in self.txt2bags:
bags = self.txt2bags[txt]
else:
bags = ST.retrieve_feature(
txt, feature_extract_config=self._feature_extract_config)
if cache:
self.txt2bags[txt] = bags
return bags
def _cross_train(self, fold_sz):
rid2shard = ST.random_shardlize(fold_sz, len(self._train_xs), load=True)
precision = 0
for fid,sd in rid2shard.items():
tmp_train_xs = [self._train_xs[i] for i in sd]
tmp_train_ys = [self._train_ys[i] for i in sd]
test_set = [(self._feature_encoding(self._train_xs[i]), self._train_ys[i]) for i in sd]
classifier = self._train(tmp_train_xs, tmp_train_ys)
p = classify.accuracy(classifier, test_set)
linfo('maxent classifier precision: %.4f' % p)
precision += p
linfo('average maxent classifier precision: %.4f' % precision/fold_sz)
def _all_train(self, total_word2cnt, total_tag2cnt):
if os.path.exists(self._test_path):
test_xs, test_ys = ST.load_data(self._test_path)
#linfo('load manually tagged data count: %s' % len(test_xs))
else:
return
test_t2c = {"P":0,"N":0,"O":0}
for y in test_ys:
if y not in test_t2c:
raise Exception('Key Error in tag2cnt. unknown key: %s' % y)
test_t2c[y] += 1
#print test_t2c
return self._batch_predict(test_xs, test_ys, total_word2cnt, total_tag2cnt, test_t2c)
def _all_train(self, total_word2cnt, total_tag2cnt):
if os.path.exists(self._test_path):
test_xs, test_ys = ST.load_data(self._test_path)
#linfo('load manually tagged data count: %s' % len(test_xs))
else:
return
test_t2c = {"P": 0, "N": 0, "O": 0}
for y in test_ys:
if y not in test_t2c:
raise Exception('Key Error in tag2cnt. unknown key: %s' % y)
test_t2c[y] += 1
#print test_t2c
return self._batch_predict(test_xs, test_ys, total_word2cnt,
total_tag2cnt, test_t2c)
def parse_topics_realtime(self):
topic_cnt, total_cnt = 0, 0
topic2txt = {}
for name, txts in self.stats:
for txt in txts:
total_cnt += 1
topic = ST.parse_topic(txt)
if not topic:
continue
topic_cnt += 1
topic2txt.setdefault(topic, list())
topic2txt[topic].append(txt)
print 'total cnt: %s. topic stats cnt: %s' % (total_cnt, topic_cnt)
print 'topic cnt: %s' % len(topic2txt)
return topic2txt
def _train(self, shard_sz=10, cross_validation=True):
#print self._ngrams_config
linfo('begin train classifier')
st = time.time()
rid2shard = ST.random_shardlize(shard_sz,
len(self._train_xs),
load=True,
path=self.rand_path)
#rid2word_info = {}
#total_word2cnt = BayesClassifier.Word2Cnt()
rid2tag_cnt, rid2word_presence = {}, {}
total_word2presence = BayesClassifier.Word2Cnt()
total_tag2cnt = {"P": 0, "N": 0, "O": 0}
for rid in range(1, shard_sz + 1):
shard = rid2shard[rid]
#rid2word_info[rid]
rid2tag_cnt[rid], rid2word_presence[rid] = self._cal_shard2info(
shard)
#for tag, w2c in rid2word_info[rid].items():
# for w, c in w2c.items():
# total_word2cnt[tag].setdefault(w, 0)
# total_word2cnt[tag][w] += c
for tag, w2p in rid2word_presence[rid].items():
for w, c in w2p.items():
total_word2presence[tag].setdefault(w, 0)
total_word2presence[tag][w] += c
for tag, cnt in rid2tag_cnt[rid].items():
total_tag2cnt[tag] += cnt
#self._debug_bigram(total_word2presence)
self._prune(total_word2presence, rid2word_presence, total_tag2cnt)
self.total_w2c, self.total_t2c = total_word2presence, total_tag2cnt
linfo(self.total_t2c)
#cross_validation
if cross_validation:
linfo('beign cross validation')
p, r, f = self._cross_train(total_word2presence, rid2word_presence,
total_tag2cnt, rid2tag_cnt, shard_sz,
rid2shard)
linfo(
'Classifier METRIC trained-precision: %.4f. recall: %.4f.f-value:%.4f. train cost used: %.2f'
% (p, r, f, time.time() - st))
else:
linfo('beign train and test with manually tagged data set')
p, r, f = self._all_train(total_word2presence, total_tag2cnt)
linfo(
'Manually Tag Data Classifier METRIC trained-precision: %.4f. recall: %.4f.f-value:%.4f. train cost used: %.2f'
% (p, r, f, time.time() - st))
def run(self, sample_enabled=False, profile_enabled=False):
for csf, path in self.classifiers:
csf.train()
action_day = ET.format_time(time.localtime())[:10]
action_total_cnt = 0
if profile_enabled:
self.reset_profile()
while True:
try:
stats = self.psr.online_run(interval=10)
if not stats:
continue
linfo('-------roundly analysis----------')
citys = map(lambda x:x[0], stats)
stats = map(lambda x:x[1], stats)
raw_stats, stats = ST.preprocess(stats)
valid_ids = [i for i, txt in enumerate(stats) if txt]
stats = map(lambda i:stats[i], valid_ids)
raw_stats = map(lambda i:raw_stats[i], valid_ids)
citys = map(lambda i:citys[i], valid_ids)
f_t = ET.format_time(time.localtime())
if sample_enabled:
sample_path = '%srealtime_%s' % (sample_prefix, f_t.replace(' ', '').replace('-', '').replace(':',''))
ET.write_file(sample_path, 'a', '%s\n' % json.dumps(raw_stats[:300]))
#only one model supported at the same time now.
for clf, path in self.classifiers:
tag2cnt = {'P':0, 'N':0, 'O':0}
pred_tags = clf.predict(stats)
for tag in pred_tags:
tag2cnt[tag] += 1
tag2dist = {tag:cnt * 1.0 / len(stats) for tag,cnt in tag2cnt.items()}
linfo('%s-roundly online sentiment distribution: %s' % (clf, tag2dist))
f_time = ET.format_time(time.localtime())
today = f_time[:10]
action_total_cnt = (action_total_cnt + len(pred_tags)) if today == action_day else len(pred_tags)
save(path, 'a', '%s\t%s\t%s\n' % (f_time, json.dumps(tag2dist), len(stats)))
if profile_enabled:
self.update_profile_spatial(citys, pred_tags)
self.update_profile_topic(raw_stats, pred_tags)
if today != action_day:
self.save_profile(action_day)
self.reset_profile()
action_day = today
break
except Exception as e:
lexcept('Unknown exception %s' % e)
def train(self,cross_validation=False, fold_sz=10, test_path='../../test_data/tri_test_data'):
self._train_xs, self._train_ys = ST.load_data(self._path)
if not self._emoticon:
ST.remove_emoticon(self._train_xs)
self.gram2gid = self._discretize_gram2gid()
if cross_validation:
linfo('begin to cross train')
self._cross_train(fold_sz)
else:
classifier = self._train(self._train_xs, self._train_ys)
self._test_xs, self._test_ys = ST.load_data(test_path)
ST.replace_url(self._test_xs, fill='H')
ST.replace_target(self._test_xs, fill='T')
test_set = [(self._feature_encoding(txt), tag) for txt, tag in zip(self._test_xs, self._test_ys)]
linfo('maxent classifier precision: %.4f' % classify.accuracy(classifier, test_set))
def __init__(self, ct='tri', prefix=''):
if prefix and prefix != 'Dg_':
raise Exception('INVALID PREFIX GIVEN!!!')
self.flag_prefix = prefix
self.train_data_path = '%s/train_data/%s%s_train_data' % (project_dir, prefix, ct)
if ct not in ['bi', 'tri']:
raise Exception('INVALID Classifier Type')
self.classifier_type = ct
self.tag2index = TAG2INDEX
self._train_xs, self._train_ys = ST.load_data(self.train_data_path)
self._train_ys = map(lambda x: self.tag2index[x], self._train_ys)
#self._feature_extract_config = ['unigram', 'bigram']
self._feature_extract_config = feature_config
linfo('feature extract config: %s' % self._feature_extract_config)
linfo('classifier type %s' % ct)
linfo('init %s success' % self)
def _cal_shard2info(self, shard_indexs):
#word2cnt = BayesClassifier.Word2Cnt()
word2presence = BayesClassifier.Word2Cnt()
#word_total_cnt = 0
tag2cnt = {"P":0,"N":0,"O":0}
for index in shard_indexs:
#word_total_cnt += len(x)
txt = self._train_xs[index]
tag = self._train_ys[index]
tag2cnt[tag] += 1
bags = ST.retrieve_feature(txt, feature_extract_config=self._ngrams_config)
for w in bags:
word2presence[tag].setdefault(w, 0)
word2presence[tag][w] += 1
#word2cnt[tag].setdefault(w, 0)
#word2cnt[tag][w] += 1
return tag2cnt, word2presence
def __init__(self, ct='tri', prefix=''):
if prefix and prefix != 'Dg_':
raise Exception('INVALID PREFIX GIVEN!!!')
self.flag_prefix = prefix
self.train_data_path = '%s/train_data/%s%s_train_data' % (project_dir,
prefix, ct)
if ct not in ['bi', 'tri']:
raise Exception('INVALID Classifier Type')
self.classifier_type = ct
self.tag2index = TAG2INDEX
self._train_xs, self._train_ys = ST.load_data(self.train_data_path)
self._train_ys = map(lambda x: self.tag2index[x], self._train_ys)
#self._feature_extract_config = ['unigram', 'bigram']
self._feature_extract_config = feature_config
linfo('feature extract config: %s' % self._feature_extract_config)
linfo('classifier type %s' % ct)
linfo('init %s success' % self)
def _cal_shard2info(self, shard_indexs):
#word2cnt = BayesClassifier.Word2Cnt()
word2presence = BayesClassifier.Word2Cnt()
#word_total_cnt = 0
tag2cnt = {"P": 0, "N": 0, "O": 0}
for index in shard_indexs:
#word_total_cnt += len(x)
txt = self._train_xs[index]
tag = self._train_ys[index]
tag2cnt[tag] += 1
bags = ST.retrieve_feature(
txt, feature_extract_config=self._ngrams_config)
for w in bags:
word2presence[tag].setdefault(w, 0)
word2presence[tag][w] += 1
#word2cnt[tag].setdefault(w, 0)
#word2cnt[tag][w] += 1
return tag2cnt, word2presence
def build_sparse_X(self, _xs):
row_num = len(_xs)
col_num = len(self.gram2gid)
rows, cols = [], []
total_cnt = 0
for i, txt in enumerate(_xs):
bags = ST.retrieve_feature(
txt, feature_extract_config=self._feature_extract_config)
for w in bags:
if w in self.gram2gid:
wid = self.gram2gid[w]
rows.append(i)
cols.append(wid)
total_cnt += 1
linfo('build scipy sparse matrice. total_valid_cnt: %s' % (total_cnt))
row = np.array(rows)
col = np.array(cols)
data = np.array([1 for i in range(total_cnt)])
mtx = sparse.csr_matrix((data, (row, col)), shape=(row_num, col_num))
return mtx
def _predict(self, txt, train_w2c, train_t2c, debug=False, emoticon=True):
#if emoticon and 'emoticon' not in self._ngrams_config:
# self._ngrams_config.append('emoticon')
#elif not emoticon and 'emoticon' in self._ngrams_config:
# self._ngrams_config = filter(lambda x: x != 'emoticon', self._ngrams_config)
#grams = self._retrieve_feature(txt)
grams = ST.retrieve_feature(txt, feature_extract_config=self._ngrams_config, gram_icon_mixed=emoticon)
if debug:
linfo('begin debug case: %s' % txt)
tag2score = {"P":0,"N":0,"O":0}
for w in grams:
for tag in tag2score:
if not train_t2c[tag]:
continue
score = self._cal_likelihood(train_w2c[tag].get(w, 0), train_t2c[tag])
tag2score[tag] += score
if debug:
linfo('DEBUG probability for gram %s when given tag %s is: %.4f. gram cnt: %s.tag cnt: %s' % (w, tag, score, train_w2c[tag].get(w, 0), train_t2c[tag]))
pred_tag = sorted(tag2score.keys(), key=lambda x: tag2score[x], reverse=True)[0]
if debug:
linfo('predict tag2score: %s' % tag2score)
return pred_tag
def _train(self, shard_sz=10, cross_validation=True):
#print self._ngrams_config
linfo('begin train classifier')
st = time.time()
rid2shard = ST.random_shardlize(shard_sz, len(self._train_xs), load=True, path=self.rand_path)
#rid2word_info = {}
#total_word2cnt = BayesClassifier.Word2Cnt()
rid2tag_cnt, rid2word_presence = {}, {}
total_word2presence = BayesClassifier.Word2Cnt()
total_tag2cnt = {"P":0,"N":0,"O":0}
for rid in range(1, shard_sz+1):
shard = rid2shard[rid]
#rid2word_info[rid]
rid2tag_cnt[rid], rid2word_presence[rid] = self._cal_shard2info(shard)
#for tag, w2c in rid2word_info[rid].items():
# for w, c in w2c.items():
# total_word2cnt[tag].setdefault(w, 0)
# total_word2cnt[tag][w] += c
for tag, w2p in rid2word_presence[rid].items():
for w, c in w2p.items():
total_word2presence[tag].setdefault(w, 0)
total_word2presence[tag][w] += c
for tag, cnt in rid2tag_cnt[rid].items():
total_tag2cnt[tag] += cnt
#self._debug_bigram(total_word2presence)
self._prune(total_word2presence, rid2word_presence, total_tag2cnt)
self.total_w2c, self.total_t2c = total_word2presence, total_tag2cnt
linfo(self.total_t2c)
#cross_validation
if cross_validation:
linfo('beign cross validation')
p, r, f= self._cross_train(total_word2presence, rid2word_presence, total_tag2cnt, rid2tag_cnt, shard_sz, rid2shard)
linfo('Classifier METRIC trained-precision: %.4f. recall: %.4f.f-value:%.4f. train cost used: %.2f' % (p , r , f, time.time()- st))
else:
linfo('beign train and test with manually tagged data set')
p, r, f = self._all_train(total_word2presence, total_tag2cnt)
linfo('Manually Tag Data Classifier METRIC trained-precision: %.4f. recall: %.4f.f-value:%.4f. train cost used: %.2f' % (p , r , f , time.time()- st))
def parse_city_public_stats(in_path='../stats/train_public_stats',
out_path='../test_data/city_test_data'):
st_t = time.time()
city2txt = {}
city_stat_cnt, total_cnt = 0, 0
stat_ids = set()
txts_upperbound = 1000
with open(in_path, 'r') as f:
for line in f:
total_cnt += 1
dic = json.loads(line.strip())
if dic['id'] in stat_ids:
continue
else:
stat_ids.add(dic['id'])
city = ST.parse_spatial(dic)
if not city:
continue
city2txt.setdefault(city, list())
if len(city2txt[city]) >= txts_upperbound:
continue
city2txt[city].append(dic['text'])
locs = sorted(city2txt.keys(),
key=lambda x: len(city2txt[x]),
reverse=True)
print 'city_stat_cnt', city_stat_cnt
print 'total_cnt', total_cnt
print 'time used: %.2f' % (time.time() - st_t)
citys = sorted(city2txt.keys())
#for x in citys:
# print x, len(city2txt[x])
if os.path.exists(out_path):
os.system('rm %s' % out_path)
for x in locs:
for txt in city2txt[x]:
dic = {x: txt}
ET.write_file(out_path, 'a', '%s\n' % json.dumps(dic))