/
MultiClassTsetlinMachine.pyx
365 lines (280 loc) · 12.6 KB
/
MultiClassTsetlinMachine.pyx
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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
# Copyright (c) 2020 Ole-Christoffer Granmo
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
# This code implements a multiclass version of the Tsetlin Machine from paper arXiv:1804.01508
# https://arxiv.org/abs/1804.01508
#cython: boundscheck=False, cdivision=True, initializedcheck=False, nonecheck=False
import numpy as np
cimport numpy as np
import random
from libc.stdlib cimport rand, RAND_MAX
########################################
### The Multiclass Tsetlin Machine #####
########################################
cdef class MultiClassTsetlinMachine:
cdef int number_of_classes
cdef int number_of_clauses
cdef int number_of_features
cdef float s
cdef int number_of_states
cdef int[:,:,:] ta_state
cdef int[:] clause_count
cdef int[:,:,:] clause_sign
cdef int[:] clause_output
cdef int[:] class_sum
cdef int[:] feedback_to_clauses
cdef int threshold
cdef int boost_true_positive_feedback
# Initialization of the Tsetlin Machine
def __init__(self, number_of_classes, number_of_clauses, number_of_features, number_of_states, s, threshold, boost_true_positive_feedback = 0):
cdef int[:] target_indexes
cdef int c,i,j,m
self.number_of_classes = number_of_classes
self.number_of_clauses = number_of_clauses
self.number_of_features = number_of_features
self.number_of_states = number_of_states
self.s = s
self.threshold = threshold
self.boost_true_positive_feedback = boost_true_positive_feedback
# The state of each Tsetlin Automaton is stored here. The automata are randomly initialized to either 'number_of_states' or 'number_of_states' + 1.
self.ta_state = np.random.choice([self.number_of_states, self.number_of_states+1], size=(self.number_of_clauses, self.number_of_features, 2)).astype(dtype=np.int32)
# Data structures for keeping track of which clause refers to which class, and the sign of the clause
self.clause_count = np.zeros((self.number_of_classes,), dtype=np.int32)
self.clause_sign = np.zeros((self.number_of_classes, self.number_of_clauses/self.number_of_classes, 2), dtype=np.int32)
# Data structures for intermediate calculations (clause output, summation of votes, and feedback to clauses)
self.clause_output = np.zeros(shape=(self.number_of_clauses,), dtype=np.int32)
self.class_sum = np.zeros(shape=(self.number_of_classes,), dtype=np.int32)
self.feedback_to_clauses = np.zeros(shape=(self.number_of_clauses), dtype=np.int32)
# Set up the Tsetlin Machine structure
for i in xrange(self.number_of_classes):
for j in xrange(self.number_of_clauses / self.number_of_classes):
self.clause_sign[i,self.clause_count[i],0] = i*(self.number_of_clauses/self.number_of_classes) + j
if j % 2 == 0:
self.clause_sign[i, self.clause_count[i], 1] = 1
else:
self.clause_sign[i, self.clause_count[i], 1] = -1
self.clause_count[i] += 1
# Calculate the output of each clause using the actions of each Tsetline Automaton.
# Output is stored an internal output array.
cdef void calculate_clause_output(self, int[:] X, int predict=0):
cdef int j,k
cdef int action_include, action_include_negated
cdef int all_exclude
for j in xrange(self.number_of_clauses):
self.clause_output[j] = 1
all_exclude = 1
for k in xrange(self.number_of_features):
action_include = self.action(self.ta_state[j,k,0])
action_include_negated = self.action(self.ta_state[j,k,1])
if action_include == 1 or action_include_negated == 1:
all_exclude = 0
if (action_include == 1 and X[k] == 0) or (action_include_negated == 1 and X[k] == 1):
self.clause_output[j] = 0
break
if predict == 1 and all_exclude == 1:
self.clause_output[j] = 0
# Sum up the votes for each class (this is the multiclass version of the Tsetlin Machine)
cdef void sum_up_class_votes(self):
cdef int target_class
cdef int j
for target_class in xrange(self.number_of_classes):
self.class_sum[target_class] = 0
for j in xrange(self.clause_count[target_class]):
self.class_sum[target_class] += self.clause_output[self.clause_sign[target_class,j,0]]*self.clause_sign[target_class,j,1]
if self.class_sum[target_class] > self.threshold:
self.class_sum[target_class] = self.threshold
elif self.class_sum[target_class] < -self.threshold:
self.class_sum[target_class] = -self.threshold
########################################
### Predict Target Class for Input X ###
########################################
def predict(self, int[:] X):
cdef int target_class
cdef int max_class
cdef float max_class_sum
###############################
### Calculate Clause Output ###
###############################
self.calculate_clause_output(X, predict=1)
###########################
### Sum up Clause Votes ###
###########################
self.sum_up_class_votes()
##########################################
### Identify Class with Largest Output ###
##########################################
max_class_sum = self.class_sum[0]
max_class = 0
for target_class in xrange(1, self.number_of_classes):
if max_class_sum < self.class_sum[target_class]:
max_class_sum = self.class_sum[target_class]
max_class = target_class
return max_class
# Translates automata state to action
cdef int action(self, int state):
if state <= self.number_of_states:
return 0
else:
return 1
# Get the state of a specific automaton, indexed by clause, feature, and automaton type (include/include negated).
def get_state(self, int clause, int feature, int automaton_type):
return self.ta_state[clause,feature,automaton_type]
############################################
### Evaluate the Trained Tsetlin Machine ###
############################################
def evaluate(self, int[:,:] X, int[:] y, int number_of_examples):
cdef int l, j
cdef int errors
cdef int max_class
cdef float max_class_sum
cdef int[:] Xi
Xi = np.zeros((self.number_of_features,), dtype=np.int32)
errors = 0
for l in xrange(number_of_examples):
###############################
### Calculate Clause Output ###
###############################
for j in xrange(self.number_of_features):
Xi[j] = X[l,j]
self.calculate_clause_output(Xi, predict=1)
###########################
### Sum up Clause Votes ###
###########################
self.sum_up_class_votes()
##########################################
### Identify Class with Largest Output ###
##########################################
max_class_sum = self.class_sum[0]
max_class = 0
for target_class in xrange(1, self.number_of_classes):
if max_class_sum < self.class_sum[target_class]:
max_class_sum = self.class_sum[target_class]
max_class = target_class
if max_class != y[l]:
errors += 1
return 1.0 - 1.0 * errors / number_of_examples
##########################################
### Online Training of Tsetlin Machine ###
##########################################
# The Tsetlin Machine can be trained incrementally, one training example at a time.
# Use this method directly for online and incremental training.
cpdef void update(self, int[:] X, int target_class):
cdef int i, j
cdef int negative_target_class
cdef int action_include, action_include_negated
# Randomly pick one of the other classes, for pairwise learning of class output
negative_target_class = int(self.number_of_classes * 0.5*rand()/(RAND_MAX/2+1))
while negative_target_class == target_class:
negative_target_class = int(self.number_of_classes * 0.5*rand()/(RAND_MAX/2+1))
###############################
### Calculate Clause Output ###
###############################
self.calculate_clause_output(X)
###########################
### Sum up Clause Votes ###
###########################
self.sum_up_class_votes()
#####################################
### Calculate Feedback to Clauses ###
#####################################
# Initialize feedback to clauses
for j in xrange(self.number_of_clauses):
self.feedback_to_clauses[j] = 0
# Calculate feedback to clauses
for j in xrange(self.clause_count[target_class]):
if 1.0*rand()/RAND_MAX > (1.0/(self.threshold*2))*(self.threshold - self.class_sum[target_class]):
continue
if self.clause_sign[target_class,j,1] >= 0:
# Type I Feedback
self.feedback_to_clauses[self.clause_sign[target_class,j,0]] = 1
else:
# Type II Feedback
self.feedback_to_clauses[self.clause_sign[target_class,j,0]] = -1
for j in xrange(self.clause_count[negative_target_class]):
if 1.0*rand()/RAND_MAX > (1.0/(self.threshold*2))*(self.threshold + self.class_sum[negative_target_class]):
continue
if self.clause_sign[negative_target_class,j,1] >= 0:
# Type II Feedback
self.feedback_to_clauses[self.clause_sign[negative_target_class,j,0]] = -1
else:
# Type I Feedback
self.feedback_to_clauses[self.clause_sign[negative_target_class,j,0]] = 1
#################################
### Train Invididual Automata ###
#################################
for j in xrange(self.number_of_clauses):
if self.feedback_to_clauses[j] > 0:
####################################################
### Type I Feedback (Combats False Negatives) ###
####################################################
if self.clause_output[j] == 0:
for k in xrange(self.number_of_features):
if 1.0*rand()/RAND_MAX <= 1.0/self.s:
if self.ta_state[j,k,0] > 1:
self.ta_state[j,k,0] -= 1
if 1.0*rand()/RAND_MAX <= 1.0/self.s:
if self.ta_state[j,k,1] > 1:
self.ta_state[j,k,1] -= 1
elif self.clause_output[j] == 1:
for k in xrange(self.number_of_features):
if X[k] == 1:
if self.boost_true_positive_feedback == 1 or 1.0*rand()/RAND_MAX <= (self.s-1)/self.s:
if self.ta_state[j,k,0] < self.number_of_states*2:
self.ta_state[j,k,0] += 1
if 1.0*rand()/RAND_MAX <= 1.0/self.s:
if self.ta_state[j,k,1] > 1:
self.ta_state[j,k,1] -= 1
elif X[k] == 0:
if self.boost_true_positive_feedback == 1 or 1.0*rand()/RAND_MAX <= (self.s-1)/self.s:
if self.ta_state[j,k,1] < self.number_of_states*2:
self.ta_state[j,k,1] += 1
if 1.0*rand()/RAND_MAX <= 1.0/self.s:
if self.ta_state[j,k,0] > 1:
self.ta_state[j,k,0] -= 1
elif self.feedback_to_clauses[j] < 0:
#####################################################
### Type II Feedback (Combats False Positives) ###
#####################################################
if self.clause_output[j] == 1:
for k in xrange(self.number_of_features):
action_include = self.action(self.ta_state[j,k,0])
action_include_negated = self.action(self.ta_state[j,k,1])
if X[k] == 0:
if action_include == 0 and self.ta_state[j,k,0] < self.number_of_states*2:
self.ta_state[j,k,0] += 1
elif X[k] == 1:
if action_include_negated == 0 and self.ta_state[j,k,1] < self.number_of_states*2:
self.ta_state[j,k,1] += 1
##############################################
### Batch Mode Training of Tsetlin Machine ###
##############################################
def fit(self, int[:,:] X, int[:] y, int number_of_examples, int epochs=100):
cdef int i, j, epoch
cdef int example_id
cdef int[:] Xi
cdef int target_class
cdef long[:] random_index
Xi = np.zeros((self.number_of_features,), dtype=np.int32)
random_index = np.arange(number_of_examples)
for epoch in xrange(epochs):
np.random.shuffle(random_index)
for i in xrange(number_of_examples):
example_id = random_index[i]
target_class = y[example_id]
for j in xrange(self.number_of_features):
Xi[j] = X[example_id,j]
self.update(Xi, target_class)
return