-
Notifications
You must be signed in to change notification settings - Fork 0
/
compute_mpe.py
260 lines (196 loc) · 8.88 KB
/
compute_mpe.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
import sys
sys.path.append("pypsdd")
from pypsdd import Vtree, SddManager, PSddManager, SddNode, Inst, io
from pypsdd import UniformSmoothing, Prior
from pysdd import sdd
""" This is the main way in which SDDs should be used to compute semantic loss.
Construct an instance from a given SDD and vtree file, and then use the available
functions for computing the most probable explanation, weighted model count, or
constructing a tensorflow circuit for integrating semantic loss into a project.
"""
class CircuitMPE:
def __init__(self, vtree_filename, sdd_filename):
# Load the Sdd using pysdd
vtree = sdd.Vtree.from_file(vtree_filename)
manager = sdd.SddManager.from_vtree(vtree)
self.alpha = manager.read_sdd_file(sdd_filename.encode())
# Load the PSdd using pypsdd
vtree = Vtree.read(vtree_filename)
manager = SddManager(vtree)
alpha = io.sdd_read(sdd_filename, manager)
# Convert to psdd
pmanager = PSddManager(vtree)
# Storing psdd
self.beta = pmanager.copy_and_normalize_sdd(alpha, vtree)
def conjoin(self, other):
gamma = self.alpha.conjoin(other.alpha)
vtree = Vtree.read(vtree_file)
manager = SddManager(vtree)
# Recreate PSDD
def compute_mpe_inst(self, lit_weights, binary_encoding=True):
mpe_inst = self.beta.get_weighted_mpe(lit_weights)[1]
if binary_encoding:
# Sort by variable, but ignoring negatives
mpe_inst.sort(key=lambda x: abs(x))
return [int(x > 0) for x in mpe_inst]
else:
return mpe_inst
def weighted_model_count(self, lit_weights):
return self.beta.weighted_model_count(lit_weights)
def get_norm_ac(self, litleaves):
return self.beta.generate_normalized_ac(litleaves)
def get_tf_ac(self, litleaves):
return self.beta.generate_tf_ac(litleaves)
def get_torch_ac(self, litleaves):
return self.beta.generate_normalized_torch_ac(litleaves)
def generate_torch_ac_stable(self, litleaves):
return self.beta.generate_normalized_torch_ac_stable(litleaves)
# Mainly used for debugging purposes
def pr_inst(self, inst):
return self.beta.pr_inst(inst)
def entropy_kld(self):
import math
pmanager = PSddManager(self.vtree)
gamma = pmanager.copy_and_normalize_sdd(self.alpha, self.vtree)
prior = UniformSmoothing(1.0)
prior.initialize_psdd(gamma)
# log model_count(beta) - ent(beta)
kld = self.beta.kl_psdd(gamma)
mc = self.beta.model_count()
entropy = -kld + math.log(mc)
return entropy
def Shannon_entropy(self):
return self.beta.Shannon_entropy()
def Shannon_entropy_stable(self):
return self.beta.Shannon_entropy_stable()
def get_models(self):
return self.beta.models(self.vtree)
def iter(sdd):
if sdd.is_decision():
for p, s in sdd.elements():
for node in iter(p): yield node
for node in iter(s): yield node
yield sdd
if __name__ == '__main__':
import torch
from torch import log
torch.set_printoptions(precision=8)
# Start test interface between pysdd and pypsdd
# Read in pysdd vtree and sdd
from pysdd import sdd
pysdd_vtree = sdd.Vtree.from_file('abcd_constraint.vtree')
pysdd_manager = sdd.SddManager.from_vtree(pysdd_vtree)
pysdd_alpha = pysdd_manager.read_sdd_file('abcd_constraint.sdd'.encode())
# Read in pypsdd vtree and ensure it matches one converted from pysdd
vtree_file = Vtree.read('abcd_constraint.vtree')
vtree = io.vtree_from_pysdd(pysdd_vtree)
assert(vtree == vtree_file)
# Create manager
manager = SddManager(vtree)
# Read in pypsdd sdd and ensure it matches one converted from pysdd
alpha_file = io.sdd_read('abcd_constraint.sdd', manager)
alpha = io.sdd_from_pysdd(pysdd_alpha, manager)
assert(alpha.is_eq(alpha_file))
## End test interface between pysdd and pypsdd
#exit()
import pdb; pdb.set_trace()
# Convert to psdd
pmanager = PSddManager(vtree)
# Storing psdd
beta = pmanager.copy_and_normalize_sdd(alpha, vtree)
prior = UniformSmoothing(1.0)
prior.initialize_psdd(beta)
# An sdd for the formula (a & b) | (c & d)
c = CircuitMPE('abcd_constraint.vtree', 'abcd_constraint.sdd')
# literal weights are of the form [[-a, a], [-b, b], [-c, c], [-d, d]]
lit_weights = torch.tensor([[0.8, 0.2], [0.7, 0.3], [0.6, 0.4], [0.9, 0.1]], device=torch.cuda.current_device())
# Weighted model counts of both the normalized and unnormalized circuits
wmc = c.get_norm_ac(lit_weights)
print(c.entropy_kld())
# Test 1
# An sdd for the formula (a & b) | (c & d)
c = CircuitMPE('abcd_constraint.vtree', 'abcd_constraint.sdd')
# literal weights are of the form [[-a, a], [-b, b], [-c, c], [-d, d]]
lit_weights = torch.tensor([[0.8, 0.2], [0.7, 0.3], [0.6, 0.4], [0.9, 0.1]], device=torch.cuda.current_device())
# Weighted model counts of both the normalized and unnormalized circuits
wmc = c.get_tf_ac(lit_weights)
wmc_normalized = c.get_torch_ac(lit_weights)
# assert the wmc of the normalized and unnormalized circuits match
assert(c.get_tf_ac(lit_weights) == 0.0976)
assert(c.get_torch_ac(lit_weights) == 0.0976)
# Entropy of the probability distribution
weights = torch.tensor([0.0224, 0.0096, 0.0056, 0.0324, 0.0036, 0.0216, 0.0024], device=torch.cuda.current_device())
probs = weights/wmc
entropy = -sum([p*log(p) for p in probs])
# Circuit Entropy
circuit_entropy = c.Shannon_entropy()
print(circuit_entropy)
print(entropy)
exit()
# Assert the circuit's entropy and the entropy of the groundtruth distribution match
assert(torch.isclose(circuit_entropy, entropy))
# Check probabilities of all the models of the formula
assert(torch.isclose(c.pr_inst([-1, -2, 3, 4]), torch.tensor(0.2295), atol=1e-04))
assert(torch.isclose(c.pr_inst([-1, 2, 3, 4]), torch.tensor(0.0984), atol=1e-04))
assert(torch.isclose(c.pr_inst([1, -2, 3, 4]), torch.tensor(0.0574), atol=1e-04))
assert(torch.isclose(c.pr_inst([1, 2, -3, -4]), torch.tensor(0.3320), atol=1e-04))
assert(torch.isclose(c.pr_inst([1, 2, -3, 4]), torch.tensor(0.0369), atol=1e-04))
assert(torch.isclose(c.pr_inst([1, 2, 3, -4]), torch.tensor(0.2213), atol=1e-04))
assert(torch.isclose(c.pr_inst([1, 2, 3, 4]), torch.tensor(0.0246), atol=1e-04))
# Test 2
# An sdd for the formula true
c = CircuitMPE('abcd_constraint.vtree', 'true_constraint.sdd')
# literal weights are of the form [[-a, a], [-b, b], [-c, c], [-d, d]]
lit_weights = torch.tensor([[0.8, 0.2], [0.7, 0.3], [0.6, 0.4], [0.9, 0.1]], device=torch.cuda.current_device())
models = [[0, 0, 0, 0],
[0, 0, 0, 1],
[0, 0, 1, 0],
[0, 0, 1, 1],
[0, 1, 0, 0],
[0, 1, 0, 1],
[0, 1, 1, 0],
[0, 1, 1, 1],
[1, 0, 0, 0],
[1, 0, 0, 1],
[1, 0, 1, 0],
[1, 0, 1, 1],
[1, 1, 0, 0],
[1, 1, 0, 1],
[1, 1, 1, 0],
[1, 1, 1, 1]]
probs = []
for model in models:
prob = 1
for i, val in enumerate(model):
prob *= lit_weights[i][val]
probs += [prob]
# Weighted model counts of both the normalized and unnormalized circuits
wmc = c.get_tf_ac(lit_weights)
wmc_normalized = c.get_torch_ac(lit_weights)
assert(wmc == wmc_normalized == 1)
# Brute force entropy
entropy = -sum([p*log(p) for p in probs])
# Circuit Entropy
circuit_entropy = c.Shannon_entropy()
assert(circuit_entropy == entropy)
# Test 3
# An sdd for the formula (P | L) & (-A | P) & (-K | (A | L))
c = CircuitMPE('LKPA_constraint.vtree', 'LKPA_constraint.sdd')
# literal weights form [[-L, L], [-K, K], [-P, P], [-A, A]]
lit_weights = torch.tensor([[0.8, 0.2], [0.7, 0.3], [0.6, 0.4], [0.9, 0.1]], device=torch.cuda.current_device())
# Weighted model counts of both the normalized and unnormalized circuits
wmc = c.get_tf_ac(lit_weights)
wmc_normalized = c.get_torch_ac(lit_weights)
# assert the wmc of the normalized and unnormalized circuits match
print(c.get_tf_ac(lit_weights))
print(c.get_torch_ac(lit_weights))
assert(c.get_tf_ac(lit_weights) == 0.4216)
assert(c.get_torch_ac(lit_weights) == 0.4216)
# Entropy of the probability distribution
weights = torch.tensor([0.2016, 0.0224, 0.0096, 0.0756, 0.0504, 0.0056, 0.0324, 0.0216, 0.0024], device=torch.cuda.current_device())
probs = weights/wmc
entropy = -sum([p*log(p) for p in probs])
# Circuit Entropy
circuit_entropy = c.Shannon_entropy()
# Assert the circuit's entropy and the entropy of the groundtruth distribution match
assert(torch.isclose(circuit_entropy, entropy))