hmm 0.0.2 → 0.1.0
Sign up to get free protection for your applications and to get access to all the features.
- data/VERSION +1 -1
- data/hmm.gemspec +2 -2
- data/lib/hmm.rb +303 -9
- data/test/test_hmm.rb +48 -14
- metadata +2 -2
data/VERSION
CHANGED
@@ -1 +1 @@
|
|
1
|
-
0.0
|
1
|
+
0.1.0
|
data/hmm.gemspec
CHANGED
@@ -5,11 +5,11 @@
|
|
5
5
|
|
6
6
|
Gem::Specification.new do |s|
|
7
7
|
s.name = %q{hmm}
|
8
|
-
s.version = "0.0
|
8
|
+
s.version = "0.1.0"
|
9
9
|
|
10
10
|
s.required_rubygems_version = Gem::Requirement.new(">= 0") if s.respond_to? :required_rubygems_version=
|
11
11
|
s.authors = ["David Tresner-Kirsch"]
|
12
|
-
s.date = %q{2009-
|
12
|
+
s.date = %q{2009-12-02}
|
13
13
|
s.description = %q{This project is a Ruby gem ('hmm') for machine learning that natively implements a (somewhat) generalized Hidden Markov Model classifier.}
|
14
14
|
s.email = %q{dwkirsch@gmail.com}
|
15
15
|
s.extra_rdoc_files = [
|
data/lib/hmm.rb
CHANGED
@@ -5,11 +5,14 @@
|
|
5
5
|
# -computing token level accuracy across a list of observation sequences
|
6
6
|
# against a provided gold standard
|
7
7
|
|
8
|
-
|
9
8
|
require 'rubygems'
|
10
9
|
require 'narray'
|
11
10
|
|
11
|
+
class Array; def sum; inject( nil ) { |sum,x| sum ? sum+x : x }; end; end
|
12
|
+
|
12
13
|
class HMM
|
14
|
+
|
15
|
+
Infinity = 1.0/0
|
13
16
|
|
14
17
|
class Classifier
|
15
18
|
attr_accessor :a, :b, :pi, :o_lex, :q_lex, :debug, :train
|
@@ -47,13 +50,296 @@ class HMM
|
|
47
50
|
end
|
48
51
|
end
|
49
52
|
|
53
|
+
# smooth to allow unobserved cases
|
54
|
+
@pi += 0.1
|
55
|
+
@a += 0.1
|
56
|
+
@b += 0.1
|
57
|
+
|
50
58
|
# normalize frequencies into probabilities
|
51
59
|
@pi /= @pi.sum
|
52
60
|
@a /= @a.sum(1)
|
53
61
|
@b /= @b.sum(1)
|
62
|
+
end
|
63
|
+
|
64
|
+
def train_unsupervised2(sequences)
|
65
|
+
# for debugging ONLY
|
66
|
+
orig_sequences = sequences.clone
|
67
|
+
sequences = [sequences.sum]
|
68
|
+
|
69
|
+
# initialize model parameters if we don't already have an estimate
|
70
|
+
@pi ||= NArray.float(@q_lex.length).fill(1)/@q_lex.length
|
71
|
+
@a ||= NArray.float(@q_lex.length, @q_lex.length).fill(1)/@q_lex.length
|
72
|
+
@b ||= NArray.float(@q_lex.length, @o_lex.length).fill(1)/@q_lex.length
|
73
|
+
puts @pi.inspect, @a.inspect, @b.inspect if debug
|
74
|
+
|
75
|
+
max_iterations = 1 #1000 #kwargs.get('max_iterations', 1000)
|
76
|
+
epsilon = 1e-6 # kwargs.get('convergence_logprob', 1e-6)
|
77
|
+
|
78
|
+
max_iterations.times do |iteration|
|
79
|
+
puts "iteration ##{iteration}" #if debug
|
80
|
+
logprob = 0.0
|
81
|
+
|
82
|
+
sequences.each do |sequence|
|
83
|
+
# just in case, skip if sequence contains unrecognized tokens
|
84
|
+
next unless (sequence-o_lex).empty?
|
85
|
+
|
86
|
+
# compute forward and backward probabilities
|
87
|
+
alpha = forward_probability(sequence)
|
88
|
+
beta = backward_probability(sequence)
|
89
|
+
lpk = log_add(alpha[-1, true]) #sum of last alphas. divide by this to get probs
|
90
|
+
logprob += lpk
|
91
|
+
|
92
|
+
xi = xi(sequence)
|
93
|
+
gamma = gamma(xi)
|
94
|
+
|
95
|
+
localA = NArray.float(q_lex.length,q_lex.length)
|
96
|
+
localB = NArray.float(q_lex.length,o_lex.length)
|
97
|
+
|
98
|
+
q_lex.each_index do |i|
|
99
|
+
q_lex.each_index do |j|
|
100
|
+
numA = -Infinity
|
101
|
+
denomA = -Infinity
|
102
|
+
sequence.each_index do |t|
|
103
|
+
break if t >= sequence.length-1
|
104
|
+
numA = log_add([numA, xi[t, i, j]])
|
105
|
+
denomA = log_add([denomA, gamma[t, i]])
|
106
|
+
end
|
107
|
+
localA[i,j] = numA - denomA
|
108
|
+
end
|
109
|
+
|
110
|
+
o_lex.each_index do |k|
|
111
|
+
numB = -Infinity
|
112
|
+
denomB = -Infinity
|
113
|
+
sequence.each_index do |t|
|
114
|
+
break if t >= sequence.length-1
|
115
|
+
denomB = log_add([denomB, gamma[t, i]])
|
116
|
+
next unless k == index(sequence[t], o_lex)
|
117
|
+
numB = log_add([numB, gamma[t, i]])
|
118
|
+
end
|
119
|
+
localB[i, k] = numB - denomB
|
120
|
+
end
|
121
|
+
|
122
|
+
end
|
123
|
+
|
124
|
+
puts "LogProb: #{logprob}"
|
125
|
+
|
126
|
+
@a = localA.collect{|x| Math::E**x}
|
127
|
+
@b = localB.collect{|x| Math::E**x}
|
128
|
+
#@pi = gamma[0, true] / gamma[0, true].sum
|
129
|
+
|
130
|
+
end
|
131
|
+
end
|
132
|
+
end
|
133
|
+
|
134
|
+
|
135
|
+
def train_unsupervised(sequences, max_iterations = 10)
|
136
|
+
# initialize model parameters if we don't already have an estimate
|
137
|
+
@pi ||= NArray.float(@q_lex.length).fill(1)/@q_lex.length
|
138
|
+
@a ||= NArray.float(@q_lex.length, @q_lex.length).fill(1)/@q_lex.length
|
139
|
+
@b ||= NArray.float(@q_lex.length, @o_lex.length).fill(1)/@q_lex.length
|
140
|
+
puts @pi.inspect, @a.inspect, @b.inspect if debug
|
141
|
+
|
142
|
+
converged = false
|
143
|
+
last_logprob = 0
|
144
|
+
iteration = 0
|
145
|
+
#max_iterations = 10 #1000 #kwargs.get('max_iterations', 1000)
|
146
|
+
epsilon = 1e-6 # kwargs.get('convergence_logprob', 1e-6)
|
147
|
+
|
148
|
+
max_iterations.times do |iteration|
|
149
|
+
puts "iteration ##{iteration}" #if debug
|
150
|
+
|
151
|
+
_A_numer = NArray.float(q_lex.length,q_lex.length).fill(-Infinity)
|
152
|
+
_B_numer = NArray.float(q_lex.length, o_lex.length).fill(-Infinity)
|
153
|
+
_A_denom = NArray.float(q_lex.length).fill(-Infinity)
|
154
|
+
_B_denom = NArray.float(q_lex.length).fill(-Infinity)
|
155
|
+
_Pi = NArray.float(q_lex.length)
|
156
|
+
|
157
|
+
logprob = 0.0
|
158
|
+
|
159
|
+
#logprob = last_logprob + 1 # take this out
|
160
|
+
|
161
|
+
sequences.each do |sequence|
|
162
|
+
# just in case, skip if sequence contains unrecognized tokens
|
163
|
+
next unless (sequence-o_lex).empty?
|
164
|
+
|
165
|
+
# compute forward and backward probabilities
|
166
|
+
alpha = forward_probability(sequence)
|
167
|
+
beta = backward_probability(sequence)
|
168
|
+
lpk = log_add(alpha[-1, true]) #sum of last alphas. divide by this to get probs
|
169
|
+
logprob += lpk
|
170
|
+
|
171
|
+
local_A_numer = NArray.float(q_lex.length,q_lex.length).fill(-Infinity)
|
172
|
+
local_B_numer = NArray.float(q_lex.length, o_lex.length).fill(-Infinity)
|
173
|
+
local_A_denom = NArray.float(q_lex.length).fill(-Infinity)
|
174
|
+
local_B_denom = NArray.float(q_lex.length).fill(-Infinity)
|
175
|
+
local_Pi = NArray.float(q_lex.length)
|
176
|
+
|
177
|
+
sequence.each_with_index do |o, t|
|
178
|
+
o_next = index(sequence[t+1], o_lex) if t < sequence.length-1
|
179
|
+
|
180
|
+
q_lex.each_index do |i|
|
181
|
+
if t < sequence.length-1
|
182
|
+
q_lex.each_index do |j|
|
183
|
+
local_A_numer[i, j] = \
|
184
|
+
log_add([local_A_numer[i, j], \
|
185
|
+
alpha[t, i] + \
|
186
|
+
log(@a[i,j]) + \
|
187
|
+
log(@b[j,o_next]) + \
|
188
|
+
beta[t+1, j]])
|
189
|
+
end
|
190
|
+
local_A_denom[i] = log_add([local_A_denom[i],
|
191
|
+
alpha[t, i] + beta[t, i]])
|
192
|
+
|
193
|
+
else
|
194
|
+
local_B_denom[i] = log_add([local_A_denom[i],
|
195
|
+
alpha[t, i] + beta[t, i]])
|
196
|
+
end
|
197
|
+
local_B_numer[i, index(o,o_lex)] = log_add([local_B_numer[i, index(o, o_lex)],
|
198
|
+
alpha[t, i] + beta[t, i]])
|
199
|
+
|
200
|
+
end
|
201
|
+
|
202
|
+
puts local_A_numer.inspect if debug
|
203
|
+
|
204
|
+
q_lex.each_index do |i|
|
205
|
+
q_lex.each_index do |j|
|
206
|
+
_A_numer[i, j] = log_add([_A_numer[i, j],
|
207
|
+
local_A_numer[i, j] - lpk])
|
208
|
+
end
|
209
|
+
o_lex.each_index do |k|
|
210
|
+
_B_numer[i, k] = log_add([_B_numer[i, k], local_B_numer[i, k] - lpk])
|
211
|
+
end
|
212
|
+
_A_denom[i] = log_add([_A_denom[i], local_A_denom[i] - lpk])
|
213
|
+
_B_denom[i] = log_add([_B_denom[i], local_B_denom[i] - lpk])
|
214
|
+
end
|
215
|
+
|
216
|
+
end
|
217
|
+
|
218
|
+
puts alpha.collect{|x| Math::E**x}.inspect if debug
|
219
|
+
end
|
220
|
+
|
221
|
+
puts _A_denom.inspect if debug
|
222
|
+
|
223
|
+
q_lex.each_index do |i|
|
224
|
+
q_lex.each_index do |j|
|
225
|
+
#puts 2**(_A_numer[i,j] - _A_denom[i]), _A_numer[i,j], _A_denom[i]
|
226
|
+
@a[i, j] = Math::E**(_A_numer[i,j] - _A_denom[i])
|
227
|
+
end
|
228
|
+
o_lex.each_index do |k|
|
229
|
+
@b[i, k] = Math::E**(_B_numer[i,k] - _B_denom[i])
|
230
|
+
end
|
231
|
+
# This comment appears in NLTK:
|
232
|
+
# Rabiner says the priors don't need to be updated. I don't
|
233
|
+
# believe him. FIXME
|
234
|
+
end
|
235
|
+
|
236
|
+
|
237
|
+
if iteration > 0 and (logprob - last_logprob).abs < epsilon
|
238
|
+
puts "CONVERGED: #{(logprob - last_logprob).abs}" if debug
|
239
|
+
puts "epsilon: #{epsilon}" if debug
|
240
|
+
break
|
241
|
+
end
|
242
|
+
|
243
|
+
puts "LogProb: #{logprob}" #if debug
|
244
|
+
|
245
|
+
last_logprob = logprob
|
246
|
+
end
|
247
|
+
end
|
248
|
+
|
249
|
+
def xi(sequence)
|
250
|
+
xi = NArray.float(sequence.length-1, q_lex.length, q_lex.length)
|
251
|
+
|
252
|
+
alpha = forward_probability(sequence)
|
253
|
+
beta = backward_probability(sequence)
|
254
|
+
|
255
|
+
0.upto sequence.length-2 do |t|
|
256
|
+
denom = 0
|
257
|
+
q_lex.each_index do |i|
|
258
|
+
q_lex.each_index do |j|
|
259
|
+
x = alpha[t, i] + log(@a[i,j]) + \
|
260
|
+
log(@b[j,index(sequence[t+1], o_lex)]) + \
|
261
|
+
beta[t+1, j]
|
262
|
+
denom = log_add([denom, x])
|
263
|
+
end
|
264
|
+
end
|
265
|
+
|
266
|
+
q_lex.each_index do |i|
|
267
|
+
q_lex.each_index do |j|
|
268
|
+
numer = alpha[t, i] + log(@a[i,j]) + \
|
269
|
+
log(@b[j,index(sequence[t+1], o_lex)]) + \
|
270
|
+
beta[t+1, j]
|
271
|
+
xi[t, i, j] = numer - denom
|
272
|
+
end
|
273
|
+
end
|
274
|
+
end
|
275
|
+
|
276
|
+
puts "Xi: #{xi.inspect}" if debug
|
277
|
+
xi
|
278
|
+
end
|
279
|
+
|
280
|
+
def gamma(xi)
|
281
|
+
gamma = NArray.float(xi.shape[0], xi.shape[1]).fill(-Infinity)
|
282
|
+
|
283
|
+
0.upto gamma.shape[0] - 1 do |t|
|
284
|
+
q_lex.each_index do |i|
|
285
|
+
q_lex.each_index do |j|
|
286
|
+
gamma[t, i] = log_add([gamma[t, i], xi[t, i, j]])
|
287
|
+
end
|
288
|
+
end
|
289
|
+
end
|
290
|
+
|
291
|
+
puts "Gamma: #{gamma.inspect}" if debug
|
292
|
+
gamma
|
293
|
+
end
|
294
|
+
|
295
|
+
def forward_probability(sequence)
|
296
|
+
alpha = NArray.float(sequence.length, q_lex.length).fill(-Infinity)
|
297
|
+
|
298
|
+
alpha[0, true] = log(@pi) + log(@b[true, index(sequence.first, o_lex)])
|
299
|
+
|
300
|
+
sequence.each_with_index do |o, t|
|
301
|
+
next if t==0
|
302
|
+
q_lex.each_index do |i|
|
303
|
+
q_lex.each_index do |j|
|
304
|
+
alpha[t, i] = log_add([alpha[t, i], alpha[t-1, j]+log(@a[j, i])])
|
305
|
+
end
|
306
|
+
alpha[t, i] += log(b[i, index(o, o_lex)])
|
307
|
+
end
|
308
|
+
end
|
309
|
+
alpha
|
54
310
|
end
|
55
|
-
|
311
|
+
|
312
|
+
def log_add(values)
|
313
|
+
x = values.max
|
314
|
+
if x > -Infinity
|
315
|
+
sum_diffs = 0
|
316
|
+
values.each do |value|
|
317
|
+
sum_diffs += Math::E**(value - x)
|
318
|
+
end
|
319
|
+
return x + log(sum_diffs)
|
320
|
+
else
|
321
|
+
return x
|
322
|
+
end
|
323
|
+
end
|
324
|
+
|
325
|
+
def backward_probability(sequence)
|
326
|
+
beta = NArray.float(sequence.length, q_lex.length).fill(-Infinity)
|
327
|
+
|
328
|
+
beta[-1, true] = log(1)
|
329
|
+
|
330
|
+
(sequence.length-2).downto(0) do |t|
|
331
|
+
q_lex.each_index do |i|
|
332
|
+
q_lex.each_index do |j|
|
333
|
+
beta[t, i] = log_add([beta[t,i], log(@a[i, j]) \
|
334
|
+
+ log(@b[j, index(sequence[t+1], o_lex)]) \
|
335
|
+
+ beta[t+1, j]])
|
336
|
+
end
|
337
|
+
end
|
338
|
+
end
|
56
339
|
|
340
|
+
beta
|
341
|
+
end
|
342
|
+
|
57
343
|
def decode(o_sequence)
|
58
344
|
# Viterbi! with log probability math to avoid underflow
|
59
345
|
|
@@ -98,19 +384,27 @@ class HMM
|
|
98
384
|
|
99
385
|
# index and deindex map between labels and the ordinals of those labels.
|
100
386
|
# the ordinals map the labels to rows and columns of Pi, A, and B
|
101
|
-
def index(
|
102
|
-
|
103
|
-
|
387
|
+
def index(subject, lexicon)
|
388
|
+
if subject.is_a?(Array) or subject.is_a?(NArray)
|
389
|
+
return subject.collect{|x| lexicon.rindex(x)}
|
390
|
+
else
|
391
|
+
return index(Array[subject], lexicon)[0]
|
392
|
+
end
|
104
393
|
end
|
105
394
|
|
395
|
+
#private
|
396
|
+
|
106
397
|
def deindex(sequence, lexicon)
|
107
398
|
sequence.collect{|i| lexicon[i]}
|
108
399
|
end
|
109
400
|
|
110
401
|
# abstracting out some array element operations for readability
|
111
|
-
def log(
|
112
|
-
|
113
|
-
|
402
|
+
def log(subject)
|
403
|
+
if subject.is_a?(Array) or subject.is_a?(NArray)
|
404
|
+
return subject.collect{|n| NMath::log n}
|
405
|
+
else
|
406
|
+
return log(Array[subject])[0]
|
407
|
+
end
|
114
408
|
end
|
115
409
|
|
116
410
|
def exp(array)
|
@@ -132,4 +426,4 @@ class HMM
|
|
132
426
|
@o, @q = o, q
|
133
427
|
end
|
134
428
|
end
|
135
|
-
end
|
429
|
+
end
|
data/test/test_hmm.rb
CHANGED
@@ -1,28 +1,62 @@
|
|
1
1
|
require 'helper'
|
2
|
+
require 'narray'
|
2
3
|
|
3
4
|
class TestHmm < Test::Unit::TestCase
|
4
|
-
|
5
|
-
|
6
|
-
|
7
|
-
end
|
8
|
-
|
9
|
-
should "decode using hand-built model" do
|
10
|
-
model = HMM::Classifier.new
|
11
|
-
|
5
|
+
def setup
|
6
|
+
@simple_model = HMM::Classifier.new
|
7
|
+
|
12
8
|
# manually build a classifier
|
13
|
-
|
14
|
-
|
15
|
-
|
9
|
+
@simple_model.o_lex = ["A", "B"]
|
10
|
+
@simple_model.q_lex = ["X", "Y", "Z"]
|
11
|
+
@simple_model.a = NArray[[0.8, 0.1, 0.1],
|
16
12
|
[0.2, 0.5, 0.3],
|
17
13
|
[0.9, 0.1, 0.0]].transpose(1,0)
|
18
|
-
|
14
|
+
@simple_model.b = NArray[ [0.2, 0.8],
|
19
15
|
[0.7, 0.3],
|
20
16
|
[0.9, 0.1]].transpose(1,0)
|
21
|
-
|
17
|
+
@simple_model.pi = NArray[0.5, 0.3, 0.2]
|
22
18
|
|
19
|
+
end
|
20
|
+
|
21
|
+
should "create new classifier" do
|
22
|
+
model = HMM::Classifier.new
|
23
|
+
assert model.class == HMM::Classifier
|
24
|
+
end
|
25
|
+
|
26
|
+
should "decode using hand-built model" do
|
23
27
|
# apply classifier to a sample observation string
|
24
|
-
q_star =
|
28
|
+
q_star = @simple_model.decode(["A","B","A"])
|
25
29
|
assert q_star == ["Z", "X", "X"]
|
26
30
|
end
|
27
31
|
|
32
|
+
should "compute forward probabilities" do
|
33
|
+
expected_alpha = NArray[ [ 0.1, 0.2272, 0.039262 ],
|
34
|
+
[ 0.21, 0.0399, 0.03038 ],
|
35
|
+
[ 0.18, 0.0073, 0.031221 ] ]
|
36
|
+
|
37
|
+
assert close_enough(expected_alpha, \
|
38
|
+
@simple_model.forward_probability(["A","B","A"]).collect{|x| Math::E**x})
|
39
|
+
end
|
40
|
+
|
41
|
+
should "compute backward probabilities" do
|
42
|
+
expected_beta = NArray[ [ 0.2271, 0.32, 1.0 ],
|
43
|
+
[ 0.1577, 0.66, 1.0 ],
|
44
|
+
[ 0.2502, 0.25, 1.0 ] ]
|
45
|
+
|
46
|
+
assert close_enough(expected_beta, \
|
47
|
+
@simple_model.backward_probability(["A","B","A"]).collect{|x| Math::E**x})
|
48
|
+
end
|
49
|
+
|
50
|
+
should "compute xi" do
|
51
|
+
@simple_model.gamma(@simple_model.xi(["A","B","A"]))
|
52
|
+
end
|
53
|
+
|
54
|
+
|
55
|
+
|
56
|
+
def close_enough(a, b)
|
57
|
+
# since we're dealing with some irrational values from logs, some checks
|
58
|
+
# need to be "good enough" rather than a perfect ==
|
59
|
+
(a-b).abs < 1e-10
|
60
|
+
end
|
61
|
+
|
28
62
|
end
|
metadata
CHANGED
@@ -1,7 +1,7 @@
|
|
1
1
|
--- !ruby/object:Gem::Specification
|
2
2
|
name: hmm
|
3
3
|
version: !ruby/object:Gem::Version
|
4
|
-
version: 0.0
|
4
|
+
version: 0.1.0
|
5
5
|
platform: ruby
|
6
6
|
authors:
|
7
7
|
- David Tresner-Kirsch
|
@@ -9,7 +9,7 @@ autorequire:
|
|
9
9
|
bindir: bin
|
10
10
|
cert_chain: []
|
11
11
|
|
12
|
-
date: 2009-
|
12
|
+
date: 2009-12-02 00:00:00 -05:00
|
13
13
|
default_executable:
|
14
14
|
dependencies:
|
15
15
|
- !ruby/object:Gem::Dependency
|