-
Notifications
You must be signed in to change notification settings - Fork 0
/
Viterbi.py
222 lines (204 loc) · 9.58 KB
/
Viterbi.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
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
import multiprocessing
import numpy as np
from itertools import product
import time
from History import History
from MLE import MLE
class Viterbi:
mle = None
tags = None
v = None
tagsToIdxDict = None
idxToTagsDict = None
seenWordsToTagsDict = None
def __init__(self, mle: MLE, allTags, v, seenWordsToTagsDict) -> None:
super().__init__()
self.k = None
self.mle = mle
self.tags = allTags
self.v = v
self.tagsToIdxDict = {}
self.idxToTagsDict = {}
self.tagsNum = len(self.tags)
self.seenWordsToTagsDict = seenWordsToTagsDict
for tag, idx in zip(self.tags, range(0, self.tagsNum)):
self.tagsToIdxDict[tag] = idx
self.idxToTagsDict[idx] = tag
self.tagsToIdxDict['*'] = self.tagsNum
self.idxToTagsDict[self.tagsNum] = '*'
input = [self.tags, self.tags]
self.relevantTagTuples = list(product(*input))
def inference(self, sentence):
self.inferenceSetUp(sentence)
self.inferenceFirstIteration(sentence)
if len(sentence) > 1:
self.inferenceSecondIteration(sentence)
for self.k in range(3, len(sentence) + 1):
self.makeNumericallyStable(self.k-1)
self.viterbiLoop()
tagsList = self.inferenceLastIteration(sentence)
return tagsList
def inferenceMP(self, sentence):
self.inferenceSetUp(sentence)
self.inferenceFirstIteration(sentence)
if len(sentence) > 1:
self.inferenceSecondIteration(sentence)
poolSize = 2
pool = multiprocessing.Pool(poolSize)
for self.k in range(3, len(sentence) + 1):
#start = time.time()
splitted = self.slice_list(list(range(0, len(self.relevantTagTuples))), poolSize)
splitted = list(filter(lambda x: len(x) > 0, splitted))
se = [(l[0], l[-1]) for l in splitted]
res = pool.imap(self.viterbiLoopMP, se)
x = np.array([np.array(x) for x in res if not x is None])
# work...
localPi = np.array([np.array(xx[0]) for xx in x])
localBp = np.array([np.array(xx[1]) for xx in x])
localPi = np.sum(localPi, axis=0)
localBp = np.sum(localBp, axis=0)
alpha = np.sum(localPi)
print(alpha,self.k)
self.pi[self.k] = localPi #/ np.sum(localPi)
self.bp[self.k] = localBp
#stop = time.time()
#print("k= ", self.k, " took ", stop - start, " seconds")
pool.close()
pool.join()
tagsList = self.inferenceLastIteration(sentence)
return tagsList
def inferenceLastIteration(self, sentence):
p = self.pi[len(sentence)]
t1, t = np.unravel_index(p.argmax(), p.shape)
tagsList = [t, t1]
tk1, tk2 = t1, t
loopSize = len(sentence) - 1
for k in reversed(range(1, loopSize)):
tk = self.bp[k + 2, tk1, tk2]
tagsList = tagsList + [tk]
tk1, tk2 = tk, tk1
tagsList = list(map(lambda x: self.idxToTagsDict[x], tagsList))
tagsList.reverse()
return tagsList
def inferenceSecondIteration(self, sentence):
for tagU, tagV in self.relevantTagTuples:
if sentence[0] in self.seenWordsToTagsDict:
if tagU not in self.seenWordsToTagsDict[sentence[0]]:
continue
if sentence[1] in self.seenWordsToTagsDict:
if tagV not in self.seenWordsToTagsDict[sentence[1]]:
continue
history = History('*', tagU, sentence, 1)
self.pi[2, self.tagsToIdxDict[tagU], self.tagsToIdxDict[tagV]] = \
self.pi[1, self.tagsToIdxDict['*'], self.tagsToIdxDict[tagU]] \
* self.mle.p(history, tagV, self.v)
self.bp[2, self.tagsToIdxDict[tagU], self.tagsToIdxDict[tagV]] = self.tagsToIdxDict['*']
def inferenceFirstIteration(self, sentence):
for tagV in self.tags:
if sentence[0] in self.seenWordsToTagsDict:
if tagV not in self.seenWordsToTagsDict[sentence[0]]:
continue
history = History('*', '*', sentence, 0)
self.pi[1, self.tagsToIdxDict['*'], self.tagsToIdxDict[tagV]] = \
self.mle.p(history, tagV, self.v)
self.bp[1, self.tagsToIdxDict['*'], self.tagsToIdxDict[tagV]] = self.tagsToIdxDict['*']
def inferenceSetUp(self, sentence):
self.sentence = sentence
self.pi = np.empty((len(sentence) + 1, len(self.tagsToIdxDict), len(self.tagsToIdxDict)))
self.bp = np.empty((len(sentence) + 1, len(self.tagsToIdxDict), len(self.tagsToIdxDict)), dtype=int)
self.pi[:, :, :] = 0
self.bp[:, :, :] = 0
self.pi[0, len(self.tagsToIdxDict) - 1, len(self.tagsToIdxDict) - 1] = 1
self.bp[0, len(self.tagsToIdxDict) - 1, len(self.tagsToIdxDict) - 1] = self.tagsNum
def viterbiLoopMP(self, se):
allTagsList, bp, k, pi, sentence = self.relevantTagTuples, self.bp, self.k, self.pi, self.sentence
myPi = np.zeros((len(self.tagsToIdxDict), len(self.tagsToIdxDict)))
myBp = np.zeros((len(self.tagsToIdxDict), len(self.tagsToIdxDict)), dtype=int)
for tagU, tagV in allTagsList[se[0]:se[1] + 1]:
if sentence[k - 1] in self.seenWordsToTagsDict:
if tagV not in self.seenWordsToTagsDict[sentence[k - 1]]:
continue
if sentence[k - 2] in self.seenWordsToTagsDict:
if tagU not in self.seenWordsToTagsDict[sentence[k - 2]]:
continue
tmpMax = -1
tmpMaxT = self.tagsNum
for tagT in self.tags:
if sentence[k - 3] in self.seenWordsToTagsDict:
t2Tags = self.seenWordsToTagsDict[sentence[k - 3]]
if tagT not in t2Tags:
continue
history = History(tagT, tagU, sentence, k - 1)
currTags = self.tags
if sentence[k-1] in self.seenWordsToTagsDict:
currTags = self.seenWordsToTagsDict[sentence[k-1]]
mleRes = self.mle.p_forTags(history, tagV, self.v, currTags)
# mleRes = self.mle.p(history, tagV, self.v)
tmpRes = pi[k - 1, self.tagsToIdxDict[tagT], self.tagsToIdxDict[tagU]] * mleRes
if tmpRes > tmpMax:
tmpMax, tmpMaxT = tmpRes, tagT
if tmpMax == -1:
for tagT in self.tags:
history = History(tagT, tagU, sentence, k - 1)
mleRes = self.mle.p(history, tagV, self.v)
tmpRes = pi[k - 1, self.tagsToIdxDict[tagT], self.tagsToIdxDict[tagU]] * mleRes
if tmpRes > tmpMax:
tmpMax, tmpMaxT = tmpRes, tagT
myPi[self.tagsToIdxDict[tagU], self.tagsToIdxDict[tagV]] = tmpMax
myBp[self.tagsToIdxDict[tagU], self.tagsToIdxDict[tagV]] = self.tagsToIdxDict[tmpMaxT]
return (myPi, myBp)
def viterbiLoop(self):
allTagsList, bp, k, pi, sentence = self.relevantTagTuples, self.bp, self.k, self.pi, self.sentence
for tagU, tagV in self.relevantTagTuples:
if sentence[k - 1] in self.seenWordsToTagsDict:
if tagV not in self.seenWordsToTagsDict[sentence[k - 1]]:
continue
if sentence[k - 2] in self.seenWordsToTagsDict:
if tagU not in self.seenWordsToTagsDict[sentence[k - 2]]:
continue
tmpMax = -1
tmpMaxT = self.tagsNum
for tagT in self.tags:
if sentence[k - 3] in self.seenWordsToTagsDict:
t2Tags = self.seenWordsToTagsDict[sentence[k - 3]]
if tagT not in t2Tags:
continue
history = History(tagT, tagU, sentence, k - 1)
currTags = self.tags
if sentence[k-1] in self.seenWordsToTagsDict:
currTags = self.seenWordsToTagsDict[sentence[k-1]]
#mleRes = self.mle.p(history, tagV, self.v)
mleRes = self.mle.p_forTags(history,tagV, self.v, currTags)
tmpRes = pi[k - 1, self.tagsToIdxDict[tagT], self.tagsToIdxDict[tagU]] * mleRes
if tmpRes > tmpMax:
tmpMax, tmpMaxT = tmpRes, tagT
if tmpMax == -1:
for tagT in self.tags:
history = History(tagT, tagU, sentence, k - 1)
mleRes = self.mle.p(history, tagV, self.v)
tmpRes = pi[k - 1, self.tagsToIdxDict[tagT], self.tagsToIdxDict[tagU]] * mleRes
if tmpRes > tmpMax:
tmpMax, tmpMaxT = tmpRes, tagT
self.pi[k,self.tagsToIdxDict[tagU], self.tagsToIdxDict[tagV]] = tmpMax
self.bp[k,self.tagsToIdxDict[tagU], self.tagsToIdxDict[tagV]] = self.tagsToIdxDict[tmpMaxT]
def slice_list(self, input, size):
input_size = len(input)
slice_size = input_size // size
remain = input_size % size
result = []
iterator = iter(input)
for i in range(size):
result.append([])
for j in range(slice_size):
result[i].append(next(iterator))
if remain:
result[i].append(next(iterator))
remain -= 1
return result
def makeNumericallyStable(self, k):
x = np.extract(self.pi[k] > 0, self.pi[k])
if x.size == 0:
self.pi[k] += 1
else:
alpha = np.average(x)
self.pi[k] = self.pi[k] / alpha