File size: 9,627 Bytes
e1c1753
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from dreamcoder.utilities import *
from dreamcoder.program import *
from dreamcoder.task import Task


class FrontierEntry(object):
    def __init__(
            self,
            program,
            _=None,
            logPrior=None,
            logLikelihood=None,
            logPosterior=None):
        self.logPosterior = logPrior + logLikelihood if logPosterior is None else logPosterior
        self.program = program
        self.logPrior = logPrior
        self.logLikelihood = logLikelihood

    def __repr__(self):
        return "FrontierEntry(program={self.program}, logPrior={self.logPrior}, logLikelihood={self.logLikelihood}".format(
            self=self)

    def strip_primitive_values(self):
        return FrontierEntry(program=strip_primitive_values(self.program),
                             logPrior=self.logPrior,
                             logPosterior=self.logPosterior,
                             logLikelihood=self.logLikelihood)
    def unstrip_primitive_values(self):
        return FrontierEntry(program=unstrip_primitive_values(self.program),
                             logPrior=self.logPrior,
                             logPosterior=self.logPosterior,
                             logLikelihood=self.logLikelihood)


class Frontier(object):
    def __init__(self, frontier, task):
        self.entries = frontier
        self.task = task

    def __repr__(
        self): return "Frontier(entries={self.entries}, task={self.task})".format(self=self)

    def __iter__(self): return iter(self.entries)

    def __len__(self): return len(self.entries)

    def json(self):
        return {"request": self.task.request.json(),
                "task": str(self.task),
                "programs": [{"program": str(e.program),
                              "logLikelihood": e.logLikelihood}
                             for e in self ]}

    def strip_primitive_values(self):
        return Frontier([e.strip_primitive_values() for e in self.entries ],
                        self.task)
    def unstrip_primitive_values(self):
        return Frontier([e.unstrip_primitive_values() for e in self.entries ],
                        self.task)

    DUMMYFRONTIERCOUNTER = 0

    @staticmethod
    def dummy(program, logLikelihood=0., logPrior=0., tp=None):
        """Creates a dummy frontier containing just this program"""
        if not tp:
            tp = program.infer().negateVariables()

        t = Task(
            "<dummy %d: %s>" %
            (Frontier.DUMMYFRONTIERCOUNTER,
             str(program)),
            tp,
            [])
        f = Frontier([FrontierEntry(program=program,
                                    logLikelihood=logLikelihood,
                                    logPrior=logPrior)],
                     task=t)
        Frontier.DUMMYFRONTIERCOUNTER += 1
        return f

    def marginalLikelihood(self):
        return lse([e.logPrior + e.logLikelihood for e in self])

    def temperature(self,T):
        """Divides prior by T"""
        return Frontier([ FrontierEntry(program=e.program,
                                        logPrior=e.logPrior/T,
                                        logLikelihood=e.logLikelihood)
                          for e in self],
                        task=self.task)
                                        

    def normalize(self):
        z = self.marginalLikelihood()
        newEntries = [
            FrontierEntry(
                program=e.program,
                logPrior=e.logPrior,
                logLikelihood=e.logLikelihood,
                logPosterior=e.logPrior +
                e.logLikelihood -
                z) for e in self]
        newEntries.sort(key=lambda e: e.logPosterior, reverse=True)
        return Frontier(newEntries,
                        self.task)

    def expectedProductionUses(self, g):
        """Returns a vector of the expected number of times each production was used"""
        import numpy as np

        this = g.rescoreFrontier(self).normalize()
        ps = list(sorted(g.primitives, key=str))
        features = np.zeros(len(ps))
        
        for j, p in enumerate(ps):
            for e in this:
                w = math.exp(e.logPosterior)
                features[j] += w * sum(child == p
                                       for _, child in e.program.walk() )
            if not p.isInvented: features[j] *= 0.3
        return features
            

    def removeZeroLikelihood(self):
        self.entries = [
            e for e in self.entries if e.logLikelihood != float('-inf')]
        return self

    def topK(self, k):
        if k == 0: return Frontier([], self.task)
        if k < 0: return self            
        newEntries = sorted(self.entries,
                            key=lambda e: (-e.logPosterior, str(e.program)))
        return Frontier(newEntries[:k], self.task)

    def sample(self):
        """Samples an entry from a frontier"""
        return sampleLogDistribution([(e.logLikelihood + e.logPrior, e)
                                      for e in self])

    @property
    def bestPosterior(self):
        return min(self.entries,
                   key=lambda e: (-e.logPosterior, str(e.program)))

    def replaceWithSupervised(self, g):
        assert self.task.supervision is not None
        return g.rescoreFrontier(Frontier([FrontierEntry(self.task.supervision,
                                                         logLikelihood=0., logPrior=0.)],
                                          task=self.task))

    @property
    def bestll(self):
        best = max(self.entries,
                   key=lambda e: e.logLikelihood)
        return best.logLikelihood


    @property
    def empty(self): return self.entries == []

    @staticmethod
    def makeEmpty(task):
        return Frontier([], task=task)

    def summarize(self):
        if self.empty:
            return "MISS " + self.task.name
        best = self.bestPosterior
        return "HIT %s w/ %s ; log prior = %f ; log likelihood = %f" % (
            self.task.name, best.program, best.logPrior, best.logLikelihood)

    def summarizeFull(self):
        if self.empty:
            return "MISS " + self.task.name
        return "\n".join([self.task.name] +
                         ["%f\t%s" % (e.logPosterior, e.program)
                          for e in self.normalize()])

    @staticmethod
    def describe(frontiers):
        numberOfHits = sum(not f.empty for f in frontiers)
        if numberOfHits > 0:
            averageLikelihood = sum(
                f.bestPosterior.logPrior for f in frontiers if not f.empty) / numberOfHits
        else:
            averageLikelihood = 0
        return "\n".join([f.summarize() for f in frontiers] +
                         ["Hits %d/%d tasks" % (numberOfHits, len(frontiers))] +
                         ["Average description length of a program solving a task: %f nats" % (-averageLikelihood)])

    def combine(self, other, tolerance=0.01):
        '''Takes the union of the programs in each of the frontiers'''
        assert self.task == other.task

        foundDifference = False

        x = {e.program: e for e in self}
        y = {e.program: e for e in other}
        programs = set(x.keys()) | set(y.keys())
        union = []
        for p in programs:
            if p in x:
                e1 = x[p]
                if p in y:
                    e2 = y[p]
                    if abs(e1.logPrior - e2.logPrior) > tolerance:
                        eprint(
                            "WARNING: Log priors differed during frontier combining: %f vs %f" %
                            (e1.logPrior, e2.logPrior))
                        eprint("WARNING: \tThe program is", p)
                        eprint()
                    if abs(e1.logLikelihood - e2.logLikelihood) > tolerance:
                        foundDifference = True
                        eprint(
                            "WARNING: Log likelihoods deferred for %s: %f & %f" %
                            (p, e1.logLikelihood, e2.logLikelihood))
                        if hasattr(self.task, 'BIC'):
                            eprint("\t%d examples, BIC=%f, parameterPenalty=%f, n parameters=%d, correct likelihood=%f" %
                                   (len(self.task.examples),
                                    self.task.BIC,
                                    self.task.BIC * math.log(len(self.task.examples)),
                                    substringOccurrences("REAL", str(p)),
                                    substringOccurrences("REAL", str(p)) * self.task.BIC * math.log(len(self.task.examples))))
                            e1.logLikelihood = - \
                                substringOccurrences("REAL", str(p)) * self.task.BIC * math.log(len(self.task.examples))
                            e2.logLikelihood = e1.logLikelihood

                        e1 = FrontierEntry(
                            program=e1.program,
                            logLikelihood=(
                                e1.logLikelihood +
                                e2.logLikelihood) /
                            2,
                            logPrior=e1.logPrior)
            else:
                e1 = y[p]
            union.append(e1)

        if foundDifference:
            eprint(
                "WARNING: Log likelihoods differed for the same program on the task %s.\n" %
                (self.task.name),
                "\tThis is acceptable only if the likelihood model is stochastic. Took the geometric mean of the likelihoods.")

        return Frontier(union, self.task)