-
Notifications
You must be signed in to change notification settings - Fork 3
/
cosine_similarity_test_large.py
164 lines (126 loc) · 6.06 KB
/
cosine_similarity_test_large.py
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
import lstm_char_cnn
import preprocess as pr
import tensorflow as tf
import numpy as np
import os
import csv
from tqdm import tqdm
data_util = pr.preprocess()
data_savepath = './npy/'
tensorflow_saver_path = './saver_large/'
time_depth = 35
word_length = 65
def load_voca():
char2idx = data_util.load_data(data_savepath+'char2idx.npy', data_structure='dictionary')
idx2char = data_util.load_data(data_savepath+'idx2char.npy', data_structure='dictionary')
word2idx = data_util.load_data(data_savepath+'word2idx.npy', data_structure='dictionary')
idx2word = data_util.load_data(data_savepath+'idx2word.npy', data_structure='dictionary')
return char2idx, idx2char, word2idx, idx2word
def _cosine_similarity(word1, word2):
# word1: [N, sum(Filters)]
# word2: [voca, sum(filters)]
word1 = np.array(word1)
word2 = np.array(word2)
dot = np.dot(word1, word2.T) # [N, voca]
word1_size = np.sqrt(np.sum(np.square(word1), axis=-1)) # [N]
word2_size = np.sqrt(np.sum(np.square(word2), axis=-1)) # [voca]
size = np.multiply(word1_size.reshape(-1, 1), word2_size) # [N, voca]
cosim = dot/size # [N, voca]
return cosim # [N, voca]
def _word2char_embedding(model, data_util, char2idx, word):
# word embedding
word2char_list = data_util._word2charidx(
word_list_1d=word,
char2idx_dict=char2idx,
word_length=word_length,
) # [len(word), word_length]
before_embedding, after_embedding = sess.run([model.before_embedding, model.after_embedding],
{
model.data:word2char_list
}
)
return before_embedding, after_embedding
def _top_k_cosine_similarity(current_word_embedding, voca_embedding, embedding_name, word, top_k=None, name=None):
cosim = _cosine_similarity(current_word_embedding, voca_embedding)
argsort = np.argsort(-cosim)[:, :top_k] # decreasing order
top_k_words = embedding_name[argsort] # [N, top_k]
top_k_cosim = np.array([cosim[index][i] for index, i in enumerate(argsort)]) # [N, top_k]
print(name)
for i in range(len(word)):
print('input_word:', word[i])
print(np.array(list(zip(top_k_words[i], top_k_cosim[i]))))
print()
def get_top_k_cosine_similarity(model, data_util, char2idx, word2idx, word, before_embedding_list, after_embedding_list, embedding_name, top_k=None):
# word2char embedding
current_word_before_embedding, current_word_after_embedding = _word2char_embedding(model, data_util, char2idx, word)
_top_k_cosine_similarity(current_word_before_embedding, before_embedding_list, embedding_name, word, top_k=5, name='before_highway_cosine_similarity')
_top_k_cosine_similarity(current_word_after_embedding, after_embedding_list, embedding_name, word, top_k=5, name='after_highway_cosine_similarity')
def make_word2char_embedding(model, data_util, char2idx, word2idx):
except_word_dict = {'</e>':0, '</p>':1, '<unk>':2}
embedding_name = np.array(list(word2idx.keys()))
word2char_list = data_util._word2charidx(
word_list_1d=embedding_name,
char2idx_dict=char2idx,
word_length=word_length,
) # [len(embedding_name), word_length]
batch_size = 20
before_embedding_list = []
after_embedding_list = []
for i in tqdm(range( int(np.ceil(len(word2char_list)/batch_size)) ), ncols=50):
batch = word2char_list[batch_size * i: batch_size * (i + 1)] # [batch_size, word_length]
before_embedding, after_embedding = sess.run([model.before_embedding, model.after_embedding],
{
model.data:batch
}
)
before_embedding_list.append(before_embedding)
after_embedding_list.append(after_embedding)
before_embedding_list = np.concatenate(before_embedding_list, axis=0) #[len(word), sum(filers)]
after_embedding_list = np.concatenate(after_embedding_list, axis=0) #[len(word), sum(filers)]
return before_embedding_list, after_embedding_list, embedding_name
char2idx, idx2char, word2idx, idx2word = load_voca()
# paper table2
cell_num = 650
voca_size = len(char2idx)
target_size = len(word2idx)
embedding_size = 15 # == projection size
lstm_stack = 2 # L=2
highway_stack = 2
pad_idx = char2idx['</p>']
window_size = [1,2,3,4,5,6,7]
filters = [min(i*50, 200) for i in window_size]
sess = tf.Session()
model = lstm_char_cnn.lstm_char_cnn(
sess = sess,
time_depth = time_depth,
word_length = word_length,
voca_size = voca_size,
target_size = target_size,
embedding_size = embedding_size,
cell_num = cell_num,
lstm_stack = lstm_stack,
highway_stack = highway_stack,
pad_idx = pad_idx,
window_size = window_size,
filters = filters
)
restore = 40
model.saver.restore(sess, tensorflow_saver_path+str(restore)+".ckpt")
print(tensorflow_saver_path+str(restore)+".ckpt")
# only first
before_embedding_list, after_embedding_list, embedding_name = make_word2char_embedding(model, data_util, char2idx, word2idx)
word = ['while', 'his', 'you', 'richard', 'trading', 'computer-aided', 'misinformed', 'looooook']
get_top_k_cosine_similarity(model, data_util, char2idx, word2idx, word, before_embedding_list, after_embedding_list, embedding_name, top_k=5)
#arithmetic test 당연히 안됨. 내적을 키우도록 학습하지 않았으므로.
before_king, after_king = _word2char_embedding(model, data_util, char2idx, ['king'])
before_man, after_man = _word2char_embedding(model, data_util, char2idx, ['man'])
before_woman, after_woman = _word2char_embedding(model, data_util, char2idx, ['woman'])
before_queen, after_queen = _word2char_embedding(model, data_util, char2idx, ['queen'])
before_calc = before_king - before_man + before_woman
after_calc = after_king - after_man + after_woman
_top_k_cosine_similarity(before_calc, before_embedding_list, embedding_name, word=['king-man+woman'], top_k=5, name='before_king-man+woman')
_top_k_cosine_similarity(after_calc, after_embedding_list, embedding_name, word=['king-man+woman'], top_k=5, name='after_king-man+woman')
print('before (king-man+woman) vs (queen) cosine similarity')
print(_cosine_similarity(before_calc, before_queen), '\n')
print('after (king-man+woman) vs (queen) cosine similarity')
print(_cosine_similarity(after_calc, after_queen))