-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathunigram_table.py
150 lines (123 loc) · 4.03 KB
/
unigram_table.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
import numpy as np
#from collate import Collate
from utils import round_number
from vocab import Vocab
class UnigramTable:
def __init__(self, max_size: int=100_000_000):
self.max_size = max_size
self.size = 0
self.z = 0
self.table = np.zeros(self.max_size)
def sample(self) -> int:
assert(0 < self.size)
unigram_idx = self.table[np.random.randint(0, self.size)]
return unigram_idx
def build(self, vocab: Vocab, alpha: float):
reserved_idxs = set(vocab.counter.keys())
free_idxs = vocab.free_idxs
# print(reserved_idxs | free_idxs)
# print(counters.to_numpy(reserved_idxs | free_idxs))
# print(counters)
# print(numpy2dict(counters.to_numpy(reserved_idxs | free_idxs)))
#print(f'vocab counter = {vocab.counter}')
counts = vocab.counter.to_numpy(reserved_idxs | free_idxs)
#print(f'counts = {counts}')
vocab_size = len(counts)
counts_pow = np.power(counts, alpha)
z = np.sum(counts_pow)
#print(f'z2 = {z}')
nums = self.max_size * counts_pow / z
#print(nums)
nums = np.vectorize(round_number)(nums)
sum_nums = np.sum(nums)
# print(f'nums = {nums}')
# print(f'sum_nums = {sum_nums}')
# print(f'max_size {self.max_size}')
while (self.max_size < sum_nums):
w = int(np.random.randint(0, vocab_size))
if 0 < nums[w]:
nums[w] -= 1
sum_nums -= 1
# print(
# f'nums = {nums}'
# )
# todo: hacer el sampleo
self.z = z
self.size = 0
# w = 0
# while w < vocab_size:
# i = 0
# while i < len(nums):
# self.table[i: ]
# print(f'vocab_size = {vocab_size}')
for w in range(vocab_size):
#print(w)
self.table[self.size: self.size + nums[w]] = w
self.size += nums[w]
# print(self.table)
# print(self.size)
def update(self, word_idx: int, F: float):
assert(0 <= word_idx)
assert(0.0 <= F)
self.z += F
if self.size < self.max_size:
if F.is_integer():
copies = min(int(F), self.max_size)
self.table[self.size: self.size + copies] = word_idx
else:
copies = min(round_number(F), self.max_size)
#print(copies)
self.table[self.size: self.size + copies] = word_idx
self.size += copies
else:
n = round_number((F / self.z) * self.max_size)
#print(f'n = {n}')
for _ in range(n):
table_idx = np.random.randint(0, self.max_size)
#print(f'table_idx ? {table_idx}')
self.table[table_idx] = word_idx
# col = Collate(4)
# col.vocab.add('hello')
# col.total_counts += 1
# col.vocab.add('how')
# col.total_counts += 1
# col.vocab.add('you')
# col.total_counts += 1
# col.vocab.add('are')
# col.total_counts += 1
# print("###########################")
# print(col.vocab.word2idx)
# col.vocab.add('hello')
# col.total_counts += 1
# col.vocab.add("how")
# col.total_counts += 1
# print("###########################")
# print(col.vocab.word2idx)
# print(col.vocab.counter)
# print(col.total_counts)
# col.reduce_vocab()
# print("###########################")
# print(col.vocab.word2idx)
# print(col.vocab.counter)
# print(col.vocab.first_full)
# print(col.vocab.free_idxs)
# print(col.total_counts)
# print("###########################")
# col.vocab.add("#")
# col.total_counts += 1
# col.vocab.add("?")
# col.total_counts += 1
# print(col.vocab.word2idx)
# print(col.vocab.counter)
# print(col.vocab.first_full)
# print(col.vocab.free_idxs)
# print(col.total_counts)
# col.vocab.delete(2)
# ut = UnigramTable(10)
# ut.build(col.vocab, col.vocab.counter, col.alpha)
# ut.update(2, 3.4)
# ut.update(3, 4.3)
# ut.update(4, 3.8)
# print(ut.table)
# ut.update(5, 2.2)
# print(ut.table)