-
Notifications
You must be signed in to change notification settings - Fork 1
/
factor.py
482 lines (435 loc) · 17.9 KB
/
factor.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
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
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
### The main file that was used to run experiments in the paper.
### Defines a simple factor class, a brute force exact inference
### algorithm, lifted orbit generation and orbit-jump MCMC.
from sage.all import *
import numpy.random
from collections import deque
import random
import numpy as np
import my_bliss
from my_graphs import *
import cProfile, pstats, StringIO
from test import *
import itertools
import time
def findsubsets(S,m):
return set(itertools.combinations(S, m))
def flip():
return numpy.random.binomial(1, 0.5)
def div_or_0(a, b):
if b == 0.0:
return 0
else:
return a / b
### returns a random element from a group g using product replacement
def fast_random_element(g):
stab = libgap(g)
return (libgap.PseudoRandom(stab).sage(parent=g)).cycle_tuples(singletons=True)
class FactorGraph:
### graph: a sage graph
### variables: a list of graph vertices which correspond to variables in the factor graph
### factors: a list of list of factors; each list is a single color
### potential: state -> real: a function which evaluates the potential on a particular state
### a state is a dictionary assigning variables to Boolean values
def __init__(self, graph, variables, factors, potential):
self.graph = graph
self.variables = variables
self.factors = factors
self.potential = potential
g = graph.automorphism_group(partition=[variables] + self.factors)
self.graph_aut = g
# print self.graph_aut
self.graph_aut_order = self.graph_aut.order()
### converts a variable state into a variable partition
def state_to_partition(self, state):
var_part_1 = []
var_part_2 = []
for v,val in state.iteritems():
if val == True:
var_part_1.append(v)
else:
var_part_2.append(v)
return [var_part_1, var_part_2]
def partition_to_state(self, part):
state1 = dict()
state2 = dict()
for v in part[0]:
state1[v] = True
state2[v] = False
for v in part[1]:
state1[v] = False
state2[v] = True
if len(part[0]) < len(self.variables) / 2.0:
return [state1, state2]
else:
return [state1]
### returns: a list of all possible states (as sorted tuples)
def gen_all_states(self):
def add_vertex(cur_lst, rst):
if len(rst) == 0:
return cur_lst
cur_add = rst.pop()
new_lst = []
for itm in cur_lst:
itm1 = itm.copy()
itm2 = itm.copy()
itm1[cur_add] = False
itm2[cur_add] = True
new_lst.append(itm1)
new_lst.append(itm2)
return add_vertex(new_lst, rst)
tbl = add_vertex([dict()], self.variables[:])
res = []
for itm in tbl:
l = itm.items()
l.sort()
res.append(tuple(l))
return res
### Given a group g, strips the factor generators from that group
def strip_factors(self, g):
stripped_gens = []
for gen in g.gens():
sgen = []
for cycle in gen.cycle_tuples(singletons=True):
# print "checking cycle " + str(cycle)
# check this cycle contains a factor
contains_factor = False
for factor in self.factors:
for e in factor:
# print "checking if " + str(e) + " in " + str(cycle)
if e in cycle:
contains_factor = True
if not contains_factor:
# print "no factor in " + str(cycle)
sgen.append(cycle)
stripped_gens.append(tuple(sgen))
return PermutationGroup(stripped_gens)
### compute partition via exhaustive enumeration using orbit generation
def partition(self):
# queue of colorings yet to be considered
queue = deque()
queue.append([[], self.variables])
g_order = self.graph_aut_order
# set of representative graph colorings
Z = 0.0
reps = set()
while len(queue) != 0:
c = queue.popleft()
# TODO: turn this into a single GI call by having it return both the
# automorphism group and the canonical form
gcanon, cert = my_bliss.canonical_form(self.graph,
partition=c + self.factors,
certificate=True,
return_graph=False)
gcanon = tuple(gcanon)
# convert the colors to their coloring in the canonical graph
c_canon = [[], []]
for c1 in c[0]:
c_canon[0].append(cert[c1])
for c1 in c[1]:
c_canon[1].append(cert[c1])
c_canon[0].sort()
c_canon[1].sort()
if (gcanon, (tuple(c_canon[0]), tuple(c_canon[1]))) in reps:
continue
A, orbits = self.graph.automorphism_group(partition=c+self.factors,
orbits=True,
algorithm="bliss")
orbitsz = g_order / A.order()
A = None
states = self.partition_to_state(c)
for s in states:
weight = orbitsz * self.potential(s)
Z += weight
reps.add((gcanon, (tuple(c_canon[0]), tuple(c_canon[1]))))
if len(c_canon[0]) + 1 <= len(self.variables) / 2.0:
# if we need the full automorphism group, we can compute it here.
# expand this node and add it to the queue
for o in orbits:
e = o.pop() # pick an element in the orbit arbitrarily
# check if this orbit is already true, or if it is any of the fixed colors
if e in c[1]:
# we know this is an uncolored vertex
newcolor = [c[0] + [e], [x for x in c[1] if x != e]]
queue.append(newcolor)
return Z
### brute forces the partition function by enumerating all possible states
### for testing and benchmarking purposes
def brute_force_partition(self):
states = self.gen_all_states()
# print "states: " + str(states)
state_to_idx = dict()
idx_to_state = dict()
for (idx, st) in enumerate(states):
state_to_idx[st] = idx
idx_to_state[idx] = st
Z = 0.0
for st in states:
Z += self.potential(dict(st))
return Z
### computes the transition matrix of a single step of the burnside process
def burnside_transition(self):
states = self.gen_all_states()
state_to_idx = dict()
idx_to_state = dict()
for (idx, st) in enumerate(states):
state_to_idx[st] = idx
idx_to_state[idx] = st
transition = np.zeros([len(states),len(states)])
for (idx, s) in enumerate(states):
var_part = self.state_to_partition(dict(s))
partition = var_part
stab = self.strip_factors(self.graph.automorphism_group(partition=partition + self.factors))
order = stab.order()
gelems = stab.list()
for e in gelems:
row_states = [dict()]
# build of list of states which are transitioned to
cyc_tuples = e.cycle_tuples(singletons=True)
for cyc in cyc_tuples:
new_states = []
for state in row_states:
st1 = state.copy()
st2 = state.copy()
for i in cyc:
st1[i] = True
st2[i] = False
new_states.append(st1)
new_states.append(st2)
row_states = new_states
for st in row_states:
l = st.items()
l.sort()
cur_idx = state_to_idx[tuple(l)]
transition[idx,cur_idx] += (1.0 / order) * (1.0 / (2**len(cyc_tuples)))
return transition
### computes the transition matrix for a Gibbs step
def gibbs_transition(self):
states = self.gen_all_states()
state_to_idx = dict()
idx_to_state = dict()
for (idx, st) in enumerate(states):
state_to_idx[st] = idx
idx_to_state[idx] = st
transition = np.zeros([len(states),len(states)])
for (idx, s) in enumerate(states):
for v in self.variables:
# compute two resulting new states
st1 = dict(s)
st2 = dict(s)
st1[v] = True
st2[v] = False
st1prob = self.potential(st1)
st1 = st1.items()
st1.sort()
st1 = tuple(st1)
st1idx = state_to_idx[st1]
st2prob = self.potential(st2)
st2 = st2.items()
st2.sort()
st2 = tuple(st2)
st2idx = state_to_idx[st2]
if st1prob + st2prob > 0:
transition[idx, st1idx] += ((1.0/ len(self.variables)) * float(st1prob)/ (st1prob + st2prob))
transition[idx, st2idx] += ((1.0 / len(self.variables)) * float(st2prob)/ (st1prob + st2prob))
else:
# stay
s_idx = state_to_idx[s]
transition[s_idx, st1idx] += 1.0 / len(self.variables)
return transition
### compute transition matrix of a within-orbit step, for implementing
### lifted MCMC
def orbit_transition(self):
states = self.gen_all_states()
state_to_idx = dict()
idx_to_state = dict()
for (idx, st) in enumerate(states):
state_to_idx[st] = idx
idx_to_state[idx] = st
transition = np.zeros([len(states),len(states)])
for (idx, s) in enumerate(states):
st_dict = dict(s)
stripped = self.strip_factors(self.graph_aut)
for g in stripped.list():
cur_s = dict()
for cyc in g.cycle_tuples(singletons=True):
for i, var in enumerate(cyc):
cur_s[var] = st_dict[cyc[(i + 1) % len(cyc)]]
st = cur_s.items()
st.sort()
st = tuple(st)
new_idx = state_to_idx[st]
transition[idx, new_idx] += 1.0 / self.graph_aut_order
return transition
### computes the metropolis hastings transition matrix for a burnside
### proposal of `burnsidesteps` steps
def burnside_mh_transition(self, burnsidesteps):
# now apply the metropolis correction for the transition x -> y, where
# each entry in `transition` contains the burnside transition
# probability
states = self.gen_all_states()
state_to_idx = dict()
idx_to_state = dict()
for (idx, st) in enumerate(states):
state_to_idx[st] = idx
idx_to_state[idx] = st
aut_cache = dict()
B = np.linalg.matrix_power(self.burnside_transition(), burnsidesteps)
for x in range(0, len(states)):
for y in range(0, len(states)):
st_x = idx_to_state[x]
part_x = self.state_to_partition(dict(st_x))
orbx = None
if st_x in aut_cache:
orbx = aut_cache[st_x]
else:
orbx = self.graph_aut_order / \
self.graph.automorphism_group(partition=self.state_to_partition(dict(st_x)) + self.factors).order()
aut_cache[st_x] = orbx
prx = self.potential(dict(st_x))
st_y = idx_to_state[y]
orby = None
if st_y in aut_cache:
orby = aut_cache[st_y]
else:
orby = self.graph_aut_order / \
self.graph.automorphism_group(partition=self.state_to_partition(dict(st_y)) + self.factors).order()
aut_cache[st_y] = orby
# orby = self.graph_aut_order / self.graph.automorphism_group(partition=self.state_to_partition(dict(st_y))).order()
pry = self.potential(dict(st_y))
if prx != 0.0:
B[x,y] = B[x,y] * np.minimum(1, (pry * orby) / (prx * orbx))
# compute probability of staying in the current position
prob_stay = 0.0
for yp in range(0, len(states)):
prob_stay += B[x,yp]
prob_stay = 1.0 - prob_stay
B[x,x] = B[x,x] + prob_stay
return B
def brute_force_prob_vector(self):
states = self.gen_all_states()
state_to_idx = dict()
idx_to_state = dict()
for (idx, st) in enumerate(states):
state_to_idx[st] = idx
idx_to_state[idx] = st
Z = 0.0
for st in states:
Z += self.potential(dict(st))
# return the vector
vec = np.zeros(len(states))
for st in states:
vec[state_to_idx[st]] = self.potential(dict(st)) / Z
return vec
### compute the total variation distance of the factor graph from its
### stationary distribution
### for each time step t, print d_tv(stationary_dist, transition_matrix^t * starting_point)
### transition_matrix: describes transition probabilities between states
### starting_point: initial probability vector on states
### num_steps: total number of steps for which to run
def total_variation(self, transition_matrix, starting_point, num_steps):
v = self.brute_force_prob_vector()
T = transition_matrix
for i in range(1, num_steps):
T = np.matmul(T, transition_matrix)
p = starting_point.dot(T)
d = 0.0
for idx in range(0, len(v)):
d += abs(v[idx] - p[idx])
print("%s\t%s" % (i, d))
### perform a single step of orbit jumping
### returns: a pair, (the ratio of transition probabilities, new state)
### n: number of burnside steps to take
def orbitjump(self, state, n):
hatx = self.burnside(state, n)
probx = self.potential(state)
probhatx = self.potential(hatx)
orbx = self.graph_aut_order / \
self.graph.automorphism_group(partition=self.state_to_partition(state) + self.factors).order()
orbhatx = self.graph_aut_order / \
self.graph.automorphism_group(partition=self.state_to_partition(hatx) + self.factors).order()
try:
transitionprob = (probhatx * orbhatx) / (probx * orbx)
return (transitionprob, hatx)
except:
# divided by 0
return (1.0, hatx)
### perform a standard orbital MCMC gibbs update, a la Niepert 2012
### state: a state
### returns: the gibbs update state
def orbitgibbs(self, state):
v = sage.misc.prandom.choice(self.variables)
# resample the variable
v_true = state.copy()
v_true[v] = True
v_false = state.copy()
v_false[v] = False
prob = None
try:
prob = self.potential(v_true) / (self.potential(v_true) + self.potential(v_false))
except:
return state # evidence was not satisfied
new_v = numpy.random.binomial(1, prob)
state[v] = new_v
# now walk along the orbit
g = fast_random_element(self.graph_aut)
# apply g to the state
for cyc in g:
fst = state[cyc[0]]
for idx, var in enumerate(cyc):
state[var] = state[cyc[(idx + 1) % len(cyc)]]
state[cyc[-1]] = fst
return state
### draw n samples according to orbit jump MCMC with no orbital MCMC
### burnsidesize: number of burnside steps to take
### n: int, total number of samples to take
### gamma: int, number of steps between taking orbit jumps
### burn: the burn in of the chain
### returns the probability of the query
def orbitjumpmcmc(self, n, query, burnsidesize=10, gamma=10, burn=100):
samples = []
# set up initial random state
cur_state = dict()
for v in self.variables:
cur_state[v] = flip()
query_count = 0.0
cur_step = 0
for i in range(0, n * burn):
if (cur_step % gamma) == 0:
# do a jump update
(ratio, new_state) = self.orbitjump(cur_state, burnsidesize)
# accept with transition probability
acceptprob = numpy.minimum(1, ratio)
if numpy.random.binomial(1, acceptprob):
cur_state = new_state
else:
# do a gibbs update
cur_state = self.orbitgibbs(cur_state)
if cur_step % burn == 0:
if query(cur_state):
query_count += 1
cur_step += 1
return query_count / n
if __name__ == "__main__":
model = gen_complete_pairwise_factorgraph_half(6)
print(model)
gibbs = model.gibbs_transition()
# print(gibbs)
# print(sum(gibbs))
within_orbit = model.orbit_transition()
orbitalmcmc = np.matmul(within_orbit, gibbs)
# unif = model.uniform_transition()
burnside = model.burnside_mh_transition(4)
M = orbitalmcmc
pv = model.brute_force_prob_vector()
start = np.zeros([2**len(model.variables)])
start[10] = 1
# print(np.linalg.matrix_power(M, 5))
print("-------------------")
print("pure gibbs")
model.total_variation(orbitalmcmc, start, 100)
print("------------------")
print("pure jump")
model.total_variation(burnside, start, 100)
# print fg.partition()
# print fg.brute_force_partition()