def input_batch_with_label(self, tw_batch, lb_batch=None):
tu.twarr_nlp(tw_batch)
self.history_length.append(len(tw_batch))
if len(self.history_length) < self.hold_batch_num:
# insufficient tweets
self.twarr.extend(tw_batch)
self.label.extend(lb_batch) if lb_batch is not None else None
return (None, None) if lb_batch is not None else None
elif not self.init_batch_ready:
# the first time when len(self.batch_twnum_list) == self.hold_batch_num, may get merged
self.init_batch_ready = True
self.twarr += tw_batch
self.label.extend(lb_batch) if lb_batch is not None else None
self.z = self.GSDPMM_twarr(list(), list(), self.twarr, iter_num=30)
z_ = [int(i) for i in self.z]
return (z_, self.label[:]) if lb_batch is not None else z_
else:
# normal process of new twarr
new_z = self.GSDPMM_twarr(self.twarr, self.z, tw_batch, iter_num=5)
self.twarr += tw_batch
self.label.extend(lb_batch) if lb_batch is not None else None
self.z += new_z
oldest_len = self.history_length.pop(0)
self.z = self.z[oldest_len:]
self.twarr = self.twarr[oldest_len:]
if lb_batch is not None:
self.label = self.label[oldest_len:]
z_ = [int(i) for i in self.z]
return (z_, self.label[:]) if lb_batch is not None else z_
def load_tw_batches(self, load_cluid_arr):
tw_batches = fu.load_array(self.labelled_batch_file)
tu.twarr_nlp(au.merge_array(tw_batches))
print("twarr nlp over")
if load_cluid_arr:
cluid_batches = fu.load_array(self.cluid_batch_file)
assert len(tw_batches) == len(cluid_batches)
for b_idx in range(len(tw_batches)):
tw_batch, cluid_batch = tw_batches[b_idx], cluid_batches[b_idx]
assert len(tw_batch) == len(cluid_batch)
for idx in range(len(tw_batch)):
tw, cluid = tw_batch[idx], cluid_batch[idx]
tw[tk.key_event_cluid] = cluid
return tw_batches
def multi(file):
# ent_tags = {'FAC', 'GPE', 'LOC', 'ORG', 'NORP'}
word_type = list()
twarr = fu.load_array(file)
twarr = tu.twarr_nlp(twarr)
for tw in twarr:
doc = tw[tk.key_spacy]
for token in doc:
word_type.append([token.text, token.ent_type_, token.tag_])
return word_type
def pre_process_twarr(twarr, lbarr):
# every tw should only get processed once
twarr = tu.twarr_nlp(twarr)
twharr = list()
for idx in range(len(twarr)):
tw, lb = twarr[idx], lbarr[idx]
twh = TweetHolder(tw)
twh.label = lb
twharr.append(twh)
return twharr
def load_tw_batches(self, load_cluid_arr):
temp_len = 60000
twarr = fu.load_array(self.filtered_twarr_file)[:temp_len]
print("load_tw_batches, len(twarr)=", len(twarr))
if load_cluid_arr:
cluidarr = fu.load_array(self.filtered_cluidarr_file)[:temp_len]
assert len(twarr) == len(cluidarr)
for idx in range(len(twarr)):
tw, twid = twarr[idx], twarr[idx][tk.key_id]
origin_id, cluid = cluidarr[idx]
assert twid == origin_id
tw[tk.key_event_cluid] = cluid
twarr = tu.twarr_nlp(twarr)
tw_batches = split_array_into_batches(twarr, self.batch_size)
print("batch distrb {}, {} batches, total {} tweets".format(
[len(b) for b in tw_batches], len(tw_batches), len(twarr)))
return tw_batches
def GSDMM_new_twarr(self, old_twarr, old_z, new_twarr, alpha, etap, etac,
etav, etah, K, iter_num):
new_twarr = tu.twarr_nlp(new_twarr)
prop_n_dict, comm_n_dict, verb_dict, ht_dict = \
self.prop_n_dict, self.comm_n_dict, self.verb_dict, self.ht_dict
D_old, D_new = len(old_twarr), len(new_twarr)
D = D_old + D_new
VP = prop_n_dict.vocabulary_size()
VC = comm_n_dict.vocabulary_size()
VV = verb_dict.vocabulary_size()
VH = ht_dict.vocabulary_size()
alpha0 = K * alpha
etap0 = VP * etap
etac0 = VC * etac
etav0 = VV * etav
etah0 = VH * etah
new_z = [-1] * D_new
m_z = [0] * K
n_z_p = [0] * K
n_z_c = [0] * K
n_z_v = [0] * K
n_z_h = [0] * K
n_zw_p = [[0] * VP for _ in range(K)]
n_zw_c = [[0] * VC for _ in range(K)]
n_zw_v = [[0] * VV for _ in range(K)]
n_zw_h = [[0] * VH for _ in range(K)]
"""initialize the counting arrays"""
def update_clu_dicts_by_tw(tw, clu_id, factor=1):
count_tw_into_tables(tw[self.key_prop_n], prop_n_dict, n_z_p,
n_zw_p, clu_id, factor)
count_tw_into_tables(tw[self.key_comm_n], comm_n_dict, n_z_c,
n_zw_c, clu_id, factor)
count_tw_into_tables(tw[self.key_verb], verb_dict, n_z_v, n_zw_v,
clu_id, factor)
count_tw_into_tables(tw[self.key_ht], ht_dict, n_z_h, n_zw_h,
clu_id, factor)
def count_tw_into_tables(tw_freq_dict_, ifd_, n_z_, n_zw_, clu_id,
factor):
for word, freq in tw_freq_dict_.word_freq_enumerate():
if factor > 0:
n_z_[clu_id] += freq
n_zw_[clu_id][ifd_.word2id(word)] += freq
else:
n_z_[clu_id] -= freq
n_zw_[clu_id][ifd_.word2id(word)] -= freq
for d in range(D_old):
k = old_z[d]
m_z[k] += 1
update_clu_dicts_by_tw(old_twarr[d], k, factor=1)
for d in range(D_new):
k = int(K * np.random.random())
new_z[d] = k
m_z[k] += 1
update_clu_dicts_by_tw(new_twarr[d], k, factor=1)
"""make sampling using current counting information"""
def rule_value_of(tw_freq_dict_, word_id_dict_, n_z_, n_zw_, p, p0,
clu_id):
i_ = value = 1.0
for word, freq in tw_freq_dict_.word_freq_enumerate():
for ii in range(0, freq):
value *= (n_zw_[clu_id][word_id_dict_.word2id(word)] + ii +
p) / (n_z_[clu_id] + i_ + p0)
i_ += 1
return value
def sample_cluster(tw, cur_iter):
prob = [0] * K
for k in range(K):
prob[k] = (m_z[k] + alpha) / (D - 1 + alpha0)
prob[k] *= rule_value_of(tw[self.key_prop_n], prop_n_dict,
n_z_p, n_zw_p, etap, etap0, k)
prob[k] *= rule_value_of(tw[self.key_comm_n], comm_n_dict,
n_z_c, n_zw_c, etac, etac0, k)
prob[k] *= rule_value_of(tw[self.key_verb], verb_dict, n_z_v,
n_zw_v, etav, etav0, k)
prob[k] *= rule_value_of(tw[self.key_ht], ht_dict, n_z_h,
n_zw_h, etah, etah0, k)
if cur_iter >= iter_num - 1:
return np.argmax(prob)
else:
return au.sample_index(np.array(prob))
"""start iteration"""
for i in range(iter_num):
for d in range(D_new):
k = new_z[d]
m_z[k] -= 1
update_clu_dicts_by_tw(new_twarr[d], k, factor=-1)
k = sample_cluster(new_twarr[d], i)
new_z[d] = k
m_z[k] += 1
update_clu_dicts_by_tw(new_twarr[d], k, factor=1)
return new_z
def pre_process_twarr(twarr):
twarr = tu.twarr_nlp(twarr)
return [TweetHolder(tw) for tw in twarr]
