MetaCat/gen.py at master · yuzhimanhua/MetaCat · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
import numpy as np
import os
np.random.seed(1234)
from spherecluster import SphericalKMeans, VonMisesFisherMixture, sample_vMF


def seed_expansion(word_sup_array, prob_sup_array, sz, write_path, vocabulary_inv, embedding_mat):
	expanded_seed = []
	vocab_sz = len(vocabulary_inv)
	for j, word_class in enumerate(word_sup_array):
		prob_sup_class = prob_sup_array[j]
		expanded_class = []
		seed_vec = np.zeros(vocab_sz)
		if len(word_class) < sz:
			for i, word in enumerate(word_class):
				seed_vec[word] = prob_sup_class[i]
			expanded = np.dot(embedding_mat.transpose(), seed_vec)
			expanded = np.dot(embedding_mat, expanded)
			word_expanded = sorted(range(len(expanded)), key=lambda k: expanded[k], reverse=True)
			for i in range(sz):
				expanded_class.append(word_expanded[i])
			expanded_seed.append(np.array(expanded_class))
		else:
			expanded_seed.append(word_class)
		if write_path is not None:
			if not os.path.exists(write_path):
				os.makedirs(write_path)
			f = open(write_path + 'class' + str(j) + '_' + str(sz) + '.txt', 'w')
			for i, word in enumerate(expanded_class):
				f.write(vocabulary_inv[word] + ' ')
			f.close()
	return expanded_seed


def label_expansion(class_labels, write_path, vocabulary_inv, embedding_mat):
	print("Retrieving top-t nearest words...")
	n_classes = len(class_labels)
	prob_sup_array = []
	current_szes = []
	all_class_labels = []
	for class_label in class_labels:
		current_sz = len(class_label)
		current_szes.append(current_sz)
		prob_sup_array.append([1/current_sz] * current_sz)
		all_class_labels += list(class_label)
	current_sz = np.min(current_szes)
	while len(all_class_labels) == len(set(all_class_labels)) and current_sz <= 200:
		current_sz += 1
		expanded_array = seed_expansion(class_labels, prob_sup_array, current_sz, None, vocabulary_inv, embedding_mat)
		all_class_labels = [w for w_class in expanded_array for w in w_class]

	expanded_array = seed_expansion(class_labels, prob_sup_array, current_sz-1, None, vocabulary_inv, embedding_mat)
	print("Final expansion size t = {}".format(len(expanded_array[0])))

	centers = []
	kappas = []
	print("Top-t nearest words for each class:")
	for i in range(n_classes):
		expanded_class = expanded_array[i]
		vocab_expanded = [vocabulary_inv[w] for w in expanded_class]
		print("Class {}:".format(i))
		print(vocab_expanded)
		expanded_mat = embedding_mat[np.asarray(expanded_class)]
		vmf_soft = VonMisesFisherMixture(n_clusters=1, n_jobs=15)
		vmf_soft.fit(expanded_mat)
		center = vmf_soft.cluster_centers_[0]
		kappa = vmf_soft.concentrations_[0]
		centers.append(center)
		kappas.append(kappa)

	for j, expanded_class in enumerate(expanded_array):
		if write_path is not None:
			if not os.path.exists(write_path):
				os.makedirs(write_path)
			f = open(write_path + 'class' + str(j) + '.txt', 'w')
			for i, word in enumerate(expanded_class):
				f.write(vocabulary_inv[word] + ' ')
			f.close()
	print("Finished vMF distribution fitting.")
	return expanded_array, centers, kappas


def pseudodocs(word_sup_array, total_num, background_array, sequence_length, len_avg,
				len_std, num_doc, interp_weight, vocabulary_inv, embedding_mat, centers, kappa, model, save_dir=None):

	for i in range(len(embedding_mat)):
		embedding_mat[i] = embedding_mat[i] / np.linalg.norm(embedding_mat[i])

	# _, centers, kappas = \
	# label_expansion(word_sup_array, save_dir, vocabulary_inv, embedding_mat)

	print("Pseudo documents generation...")
	background_vec = interp_weight * background_array
	if model == 'cnn':
		docs = np.zeros((num_doc*len(word_sup_array), sequence_length), dtype='int32')
		label = np.zeros((num_doc*len(word_sup_array), len(word_sup_array)))
		for i in range(len(word_sup_array)):
			docs_len = len_avg*np.ones(num_doc)
			center = centers[i]
			# kappa = kappas[i]
			discourses = sample_vMF(center, kappa, num_doc)
			for j in range(num_doc):
				discourse = discourses[j]
				prob_vec = np.dot(embedding_mat, discourse)
				prob_vec = np.exp(prob_vec)
				sorted_idx = np.argsort(prob_vec)[::-1]
				delete_idx = sorted_idx[total_num:]
				prob_vec[delete_idx] = 0
				prob_vec /= np.sum(prob_vec)
				prob_vec *= 1 - interp_weight
				prob_vec += background_vec
				doc_len = int(docs_len[j])
				docs[i*num_doc+j][:doc_len] = np.random.choice(len(prob_vec), size=doc_len, p=prob_vec)
				label[i*num_doc+j] = interp_weight/len(word_sup_array)*np.ones(len(word_sup_array))
				label[i*num_doc+j][i] += 1 - interp_weight
	elif model == 'rnn':
		docs = np.zeros((num_doc*len(word_sup_array), sequence_length[0], sequence_length[1]), dtype='int32')
		label = np.zeros((num_doc*len(word_sup_array), len(word_sup_array)))
		doc_len = int(len_avg[0])
		sent_len = int(len_avg[1])
		for period_idx in vocabulary_inv:
			if vocabulary_inv[period_idx] == '.':
				break
		for i in range(len(word_sup_array)):
			center = centers[i]
			# kappa = kappas[i]
			discourses = sample_vMF(center, kappa, num_doc)
			for j in range(num_doc):
				discourse = discourses[j]
				prob_vec = np.dot(embedding_mat, discourse)
				prob_vec = np.exp(prob_vec)
				sorted_idx = np.argsort(prob_vec)[::-1]
				delete_idx = sorted_idx[total_num:]
				prob_vec[delete_idx] = 0
				prob_vec /= np.sum(prob_vec)
				prob_vec *= 1 - interp_weight
				prob_vec += background_vec
				for k in range(doc_len):
					docs[i*num_doc+j][k][:sent_len] = np.random.choice(len(prob_vec), size=sent_len, p=prob_vec)
					docs[i*num_doc+j][k][sent_len] = period_idx
				label[i*num_doc+j] = interp_weight/len(word_sup_array)*np.ones(len(word_sup_array))
				label[i*num_doc+j][i] += 1 - interp_weight

	print("Finished Pseudo documents generation.")
	return docs, label


def augment(x, sup_idx, total_len):
	print("Labeled documents augmentation...")
	print(sup_idx)
	docs = x[sup_idx.flatten()]
	curr_len = len(docs)
	copy_times = int(total_len/curr_len) - 1
	y = np.zeros(len(sup_idx.flatten()), dtype='int32')
	label_nums = [len(seed_idx) for seed_idx in sup_idx]
	cnt = 0
	for i in range(len(sup_idx)):
		y[cnt:cnt+label_nums[i]] = i
		cnt += label_nums[i]

	new_docs = docs
	new_y = y
	for i in range(copy_times):
		new_docs = np.concatenate((new_docs, docs), axis=0)
		new_y = np.concatenate((new_y, y), axis=0)

	pretrain_labels = np.zeros((len(new_y),len(np.unique(y))))
	for i in range(len(new_y)):
		pretrain_labels[i][new_y[i]] = 1.0

	print("Finished labeled documents augmentation.")
	return new_docs, pretrain_labels