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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
|
import numpy
import numpy.random
import scipy.sparse
def diag_dot(a, b, out=None):
"""Input: a and b are arrays.
Output: a column vector of the dot product of the rows of a and respective
columns of b, in other words, diag(a.dot(b))."""
if out is None:
out = numpy.empty((a.shape[0], 1))
numpy.einsum('ji,ij->j', a, b, out=out[:,0])
return out
class NeuralLM(object):
def __init__(self, ngram_size, input_vocab_size, output_vocab_size, input_embedding_dimension, num_hidden, output_embedding_dimension, activation_function):
self.input_vocab_size = input_vocab_size
self.output_vocab_size = output_vocab_size
self.index_to_word = []
self.index_to_word_input = []
self.index_to_word_output = []
self.word_to_index = {}
self.word_to_index_input = {}
self.word_to_index_output = {}
self.ngram_size = ngram_size
self.input_embedding_dimension = input_embedding_dimension
self.num_hidden = num_hidden
self.output_embedding_dimension = output_embedding_dimension
self.activation_function = activation_function
self.input_embeddings = numpy.zeros((input_vocab_size, input_embedding_dimension))
if not num_hidden:
self.hidden1_weights = numpy.zeros((output_embedding_dimension, (ngram_size-1)*input_embedding_dimension))
self.hidden1_biases = numpy.zeros((output_embedding_dimension, 1))
self.hidden2_weights = numpy.zeros((1, 1))
self.hidden2_biases = numpy.zeros((1, 1))
else:
self.hidden1_weights = numpy.zeros((num_hidden, (ngram_size-1)*input_embedding_dimension))
self.hidden1_biases = numpy.zeros((num_hidden, 1))
self.hidden2_weights = numpy.zeros((output_embedding_dimension, num_hidden))
self.hidden2_biases = numpy.zeros((output_embedding_dimension, 1))
self.output_weights = numpy.zeros((output_vocab_size, output_embedding_dimension))
self.output_biases = numpy.zeros((output_vocab_size, 1))
def initialize(self, r):
def uniform(m):
m[:,:] = numpy.random.uniform(-r, r, m.shape)
uniform(self.input_embeddings)
uniform(self.hidden1_weights)
uniform(self.hidden1_biases)
uniform(self.hidden2_weights)
uniform(self.hidden2_biases)
uniform(self.output_weights)
uniform(self.output_biases)
def forward_prop(self, inputs, output=None, normalize=True):
u = numpy.bmat([[self.input_embeddings.T * ui] for ui in inputs])
h1 = numpy.maximum(0., self.hidden1_weights * u + self.hidden1_biases)
if not self.num_hidden:
h2 = h1
else:
h2 = numpy.maximum(0., self.hidden2_weights * h1 + self.hidden2_biases)
if output is None:
o = self.output_weights * h2 + self.output_biases
else:
# Inefficient version:
#o = diag_dot(output.T, (self.output_weights * h2 + self.output_biases))
#o = output.multiply(self.output_weights * h2 + self.output_biases)
# Since output is sparse, distributing multiplication by output
# is much more efficient:
o = diag_dot(output.T * self.output_weights, h2) + output.T * self.output_biases
return o
def backward_prop(self, g_output):
pass
def to_file(self, outfile):
def write_matrix(m):
for i in xrange(m.shape[0]):
outfile.write("\t".join(map(str, m[i])))
outfile.write("\n")
outfile.write("\n")
def write_vector(m):
for i in xrange(m.shape[0]):
outfile.write(str(m[i]))
outfile.write("\n")
outfile.write("\n")
outfile.write("\\config\n")
outfile.write("version 1\n")
outfile.write("ngram_size %d\n" % self.ngram_size)
outfile.write("input_vocab_size %d\n" % self.input_vocab_size)
outfile.write("output_vocab_size %d\n" % self.output_vocab_size)
outfile.write("input_embedding_dimension %d\n" % self.input_embedding_dimension)
outfile.write("num_hidden %d\n" % self.num_hidden)
outfile.write("output_embedding_dimension %d\n" % self.output_embedding_dimension)
outfile.write("activation_function %s\n" % self.activation_function)
outfile.write("\n")
if self.index_to_word_input and self.index_to_word_output:
outfile.write("\\input_vocab\n")
for word in self.index_to_word_input:
outfile.write(word + "\n")
outfile.write("\n")
outfile.write("\\output_vocab\n")
for word in self.index_to_word_output:
outfile.write(word + "\n")
outfile.write("\n")
elif self.index_to_word:
outfile.write("\\vocab\n")
for word in self.index_to_word:
outfile.write(word + "\n")
outfile.write("\n")
outfile.write("\\input_embeddings\n")
write_matrix(self.input_embeddings)
outfile.write("\\hidden_weights 1\n")
write_matrix(self.hidden1_weights)
outfile.write("\\hidden_biases 1\n")
write_matrix(self.hidden1_biases)
outfile.write("\\hidden_weights 2\n")
write_matrix(self.hidden2_weights)
outfile.write("\\hidden_biases 2\n")
write_matrix(self.hidden2_biases)
outfile.write("\\output_weights\n")
write_matrix(self.output_weights)
outfile.write("\\output_biases\n")
write_matrix(self.output_biases)
outfile.write("\\end\n")
@staticmethod
def from_file(infile):
"""Create a NeuralLM from a text file."""
# Helper functions
def read_sections(infile):
while True:
line = infile.next().strip()
if line == "\\end":
break
elif line.startswith('\\'):
yield line, read_section(infile)
def read_section(infile):
while True:
line = infile.next().strip()
if line == "":
break
else:
yield line
def read_matrix(lines, m, n, out=None):
if out is None:
out = numpy.zeros((m, n))
i = 0
for line in lines:
row = numpy.array(map(float, line.split()))
if len(row) != n:
raise Exception("wrong number of columns (expected %d, found %d)" % (n, len(row)))
if i >= m:
raise Exception("wrong number of rows (expected %d, found more)" % m)
out[i,:] = row
i += 1
if i < m:
raise Exception("wrong number of rows (expected %d, found %d)" % (m, i))
return out
if isinstance(infile, str):
infile = open(infile)
for section, lines in read_sections(infile):
if section == "\\config":
config = {}
for line in lines:
key, value = line.split()
config[key] = value
m = NeuralLM(ngram_size=int(config['ngram_size']),
input_vocab_size=int(config['input_vocab_size']),
output_vocab_size=int(config['output_vocab_size']),
input_embedding_dimension=int(config['input_embedding_dimension']),
num_hidden=int(config['num_hidden']),
output_embedding_dimension=int(config['output_embedding_dimension']),
activation_function=config['activation_function'])
elif section == "\\input_vocab":
for line in lines:
m.index_to_word_input.append(line)
m.word_to_index_input = dict((w,i) for (i,w) in enumerate(m.index_to_word_input))
elif section == "\\output_vocab":
for line in lines:
m.index_to_word_output.append(line)
m.word_to_index_output = dict((w,i) for (i,w) in enumerate(m.index_to_word_output))
elif section == "\\vocab":
for line in lines:
m.index_to_word.append(line)
m.word_to_index = dict((w,i) for (i,w) in enumerate(m.index_to_word))
elif section == "\\input_embeddings":
read_matrix(lines, m.input_vocab_size, m.input_embedding_dimension, out=m.input_embeddings)
elif section == "\\hidden_weights 1":
if not m.num_hidden:
read_matrix(lines, m.output_embedding_dimension, (m.ngram_size-1)*m.input_embedding_dimension, out=m.hidden1_weights)
else:
read_matrix(lines, m.num_hidden, (m.ngram_size-1)*m.input_embedding_dimension, out=m.hidden1_weights)
elif section == "\\hidden_biases 1":
if not m.num_hidden:
read_matrix(lines, m.output_embedding_dimension, 1, out=m.hidden1_biases)
else:
read_matrix(lines, m.num_hidden, 1, out=m.hidden1_biases)
elif section == "\\hidden_weights 2":
if not m.num_hidden:
read_matrix(lines, 1, 1, out=m.hidden2_weights)
else:
read_matrix(lines, m.output_embedding_dimension, m.num_hidden, out=m.hidden2_weights)
elif section == "\\hidden_biases 2":
if not m.num_hidden:
read_matrix(lines, 1, 1, out=m.hidden2_biases)
else:
read_matrix(lines, m.output_embedding_dimension, 1, out=m.hidden2_biases)
elif section == "\\output_weights":
read_matrix(lines, m.output_vocab_size, m.output_embedding_dimension, out=m.output_weights)
elif section == "\\output_biases":
read_matrix(lines, m.output_vocab_size, 1, out=m.output_biases)
return m
def make_data(self, ngrams):
"""Takes a list of n-grams of words (as ints),
and converts into a list of n sparse arrays."""
rows = [[] for j in xrange(self.ngram_size)]
cols = [[] for j in xrange(self.ngram_size)]
values = [[] for j in xrange(self.ngram_size)]
for i, ngram in enumerate(ngrams):
for j, w in enumerate(ngram):
rows[j].append(w)
cols[j].append(i)
values[j].append(1)
data = [scipy.sparse.csc_matrix((values[j], (rows[j], cols[j])), shape=(self.input_vocab_size, len(ngrams))) for j in xrange(self.ngram_size-1)]
data.append(scipy.sparse.csc_matrix((values[-1], (rows[-1], cols[-1])), shape=(self.output_vocab_size, len(ngrams))))
return data
|
