-
Notifications
You must be signed in to change notification settings - Fork 3
/
word2vec.lua
178 lines (139 loc) · 4.43 KB
/
word2vec.lua
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
173
174
175
176
177
require 'mobdebug'.start()
require 'nn'
require 'nngraph'
require 'optim'
require 'Embedding'
local model_utils=require 'model_utils'
require 'table_utils'
nngraph.setDebug(true)
function get_context_words(sentence, context_size, center_word_index)
local possible= {}
for i = -context_size, context_size do
if i ~= 0 and i + center_word_index <= #sentence and i + center_word_index > 0 then
possible[#possible+ 1] = sentence[i + center_word_index]
end
end
local ids = {}
for k = 1, 2*context_size do
ids[#ids + 1] = possible[math.random(1, #possible)]
end
return ids
end
function read_words(fn)
fd = io.lines(fn)
sentences = {}
line = fd()
while line do
sentence = {}
for _, word in pairs(string.split(line, " ")) do
sentence[#sentence + 1] = word
end
sentences[#sentences + 1] = sentence
line = fd()
end
return sentences
end
function math.clamp(x, min_val, max_val)
if x < min_val then
x = min_val
elseif x > max_val then
x = max_val
end
return x
end
function convert2tensors(sentences)
l = {}
for _, sentence in pairs(sentences) do
t = torch.zeros(1, #sentence)
for i = 1, #sentence do
t[1][i] = sentence[i]
end
l[#l + 1] = t
end
return l
end
sentences, vocabulary, inv_vocabulary = unpack(torch.load('filter_sentences_output.t7'))
n_data = #sentences
vocab_size = #vocabulary
function calc_max_sentence_len(sentences)
local m = 1
for _, sentence in pairs(sentences) do
m = math.max(m, #sentence)
end
return m
end
max_sentence_len = calc_max_sentence_len(sentences)
context_size = 5
batch_size = 1000
neg_samples_num = 10
data_index = 1
function gen_batch()
start_index = data_index
end_index = math.min(n_data, start_index + batch_size - 1)
if end_index == n_data then
data_index = 1
else
data_index = data_index + batch_size
end
basic_batch_size = end_index - start_index + 1
local center_words = torch.Tensor( (2*context_size * (1 + neg_samples_num)) * basic_batch_size)
local outer_words = torch.Tensor( (2*context_size * (1 + neg_samples_num)) * basic_batch_size)
local labels = torch.Tensor( center_words:size(1))
row = 1
for k = 1, basic_batch_size do
sentence = sentences[start_index + k - 1]
center_word_index = math.random(1, #sentence)
center_word = sentence[center_word_index]
context_words = get_context_words(sentence, context_size, center_word_index)
for _, outer_word in pairs(context_words) do
center_words[row] = center_word
outer_words[row] = outer_word
labels[row] = 1
row = row + 1
neg_samples = torch.rand(neg_samples_num):mul(vocab_size):floor():add(1)
outer_words[{{row, row+neg_samples_num-1}}] = neg_samples
center_words[{{row, row+neg_samples_num-1}}]:fill(center_word)
labels[{{row, row+neg_samples_num-1}}] = torch.Tensor(neg_samples_num):fill(-1)
row = row + neg_samples_num
dummy_pass = 1
end
end
return center_words, outer_words, labels
end
word_center = nn.Identity()()
word_outer = nn.Identity()()
x_center = Embedding(vocab_size, 10)(word_center)
x_outer = Embedding(vocab_size, 10)(word_outer)
z = nn.CosineDistance()({x_outer, x_center})
m = nn.gModule({word_center, word_outer}, {z, x_outer, x_center})
local params, grad_params = model_utils.combine_all_parameters(m)
params:uniform(-0.08, 0.08)
criterion = nn.MarginCriterion()
function feval(x_arg)
if x_arg ~= params then
params:copy(x_arg)
end
grad_params:zero()
local loss = 0
center_words, outer_words, labels = gen_batch()
------------------- forward pass -------------------
z, x_outer, x_center = unpack(m:forward({center_words, outer_words}))
loss_m = criterion:forward(z, labels)
loss = loss + loss_m
-- complete reverse order of the above
dx_outer = torch.zeros(x_outer:size())
dx_center = torch.zeros(x_center:size())
dz = criterion:backward(z, labels)
dcenter_words, douter_words = unpack(m:backward({center_words, outer_words}, {dz, dx_outer, dx_center}))
return loss, grad_params
end
optim_state = {learningRate = 1e-2}
for i = 1, 1000000 do
local _, loss = optim.adam(feval, params, optim_state)
if i % 1 == 0 then
print(string.format( 'loss = %6.8f, grad_params:norm() = %6.4e, params:norm() = %6.4e', loss[1], grad_params:norm(), params:norm()))
end
if i % 10 == 0 then
torch.save('model.t7', m)
end
end