File size: 19,764 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 |
from dreamcoder.fragmentUtilities import *
from dreamcoder.grammar import *
from dreamcoder.program import *
from itertools import chain
import time
class FragmentGrammar(object):
def __init__(self, logVariable, productions):
self.logVariable = logVariable
self.productions = productions
self.likelihoodCache = {}
def clearCache(self):
self.likelihoodCache = {}
def __repr__(self):
return "FragmentGrammar(logVariable={self.logVariable}, productions={self.productions}".format(
self=self)
def __str__(self):
def productionKey(xxx_todo_changeme):
(l, t, p) = xxx_todo_changeme
return not isinstance(p, Primitive), -l
return "\n".join(["%f\tt0\t$_" % self.logVariable] + ["%f\t%s\t%s" % (l, t, p)
for l, t, p in sorted(self.productions, key=productionKey)])
def buildCandidates(self, context, environment, request):
candidates = []
variableCandidates = []
for l, t, p in self.productions:
try:
newContext, t = t.instantiate(context)
newContext = newContext.unify(t.returns(), request)
candidates.append((l, newContext,
t.apply(newContext),
p))
except UnificationFailure:
continue
for j, t in enumerate(environment):
try:
newContext = context.unify(t.returns(), request)
variableCandidates.append((newContext,
t.apply(newContext),
Index(j)))
except UnificationFailure:
continue
if variableCandidates:
z = math.log(len(variableCandidates))
for newContext, newType, index in variableCandidates:
candidates.append(
(self.logVariable - z, newContext, newType, index))
z = lse([candidate[0] for candidate in candidates])
return [(l - z, c, t, p) for l, c, t, p in candidates]
def logLikelihood(self, request, expression):
_, l, _ = self._logLikelihood(Context.EMPTY, [], request, expression)
if invalid(l):
f = 'failures/likelihoodFailure%s.pickle' % (time() + getPID())
eprint("PANIC: Invalid log likelihood. expression:",
expression, "tp:", request, "Exported to:", f)
with open(f, 'wb') as handle:
pickle.dump((self, request, expression), handle)
assert False
return l
def closedUses(self, request, expression):
_, l, u = self._logLikelihood(Context.EMPTY, [], request, expression)
return l, u
def _logLikelihood(self, context, environment, request, expression):
'''returns (context, log likelihood, uses)'''
# We can cash likelihood calculations faster whenever they don't involve type inference
# This is because they are guaranteed to not modify the context,
polymorphic = request.isPolymorphic or any(
v.isPolymorphic for v in environment)
# For some reason polymorphic caching slows it down
shouldDoCaching = not polymorphic
# Caching
if shouldDoCaching:
if polymorphic:
inTypes = canonicalTypes(
[request.apply(context)] + [v.apply(context) for v in environment])
else:
inTypes = canonicalTypes([request] + environment)
cacheKey = (tuple(inTypes), expression)
if cacheKey in self.likelihoodCache:
outTypes, l, u = self.likelihoodCache[cacheKey]
context, instantiatedTypes = instantiateTypes(
context, outTypes)
outRequest = instantiatedTypes[0]
outEnvironment = instantiatedTypes[1:]
# eprint("request:", request.apply(context), "environment:",
# [ v.apply(context) for v in environment ])
# eprint("will be unified with: out request:",outRequest,"out environment",outEnvironment)
if polymorphic:
context = context.unify(request, outRequest)
for v, vp in zip(environment, outEnvironment):
context = context.unify(v, vp)
return context, l, u
if request.isArrow():
if not isinstance(expression, Abstraction):
return (context, NEGATIVEINFINITY, Uses.empty)
return self._logLikelihood(context,
[request.arguments[0]] + environment,
request.arguments[1],
expression.body)
# Not a function type
# Construct and normalize the candidate productions
candidates = self.buildCandidates(context, environment, request)
# Consider each way of breaking the expression up into a
# function and arguments
totalLikelihood = NEGATIVEINFINITY
weightedUses = []
possibleVariables = float(int(any(isinstance(candidate, Index)
for _, _, _, candidate in candidates)))
possibleUses = {candidate: 1. for _, _, _, candidate in candidates
if not isinstance(candidate, Index)}
for f, xs in expression.applicationParses():
for candidateLikelihood, newContext, tp, production in candidates:
variableBindings = {}
# This is a variable in the environment
if production.isIndex:
if production != f:
continue
else:
try:
newContext, fragmentType, variableBindings = \
Matcher.match(newContext, production, f, len(xs))
# This is necessary because the types of the variable
# bindings and holes need to match up w/ request
fragmentTypeTemplate = request
for _ in xs:
newContext, newVariable = newContext.makeVariable()
fragmentTypeTemplate = arrow(
newVariable, fragmentTypeTemplate)
newContext = newContext.unify(
fragmentType, fragmentTypeTemplate)
# update the unified type
tp = fragmentType.apply(newContext)
except MatchFailure:
continue
argumentTypes = tp.functionArguments()
if len(xs) != len(argumentTypes):
# I think that this is some kind of bug. But I can't figure it out right now.
# As a hack, count this as though it were a failure
continue
#raise GrammarFailure('len(xs) != len(argumentTypes): tp={}, xs={}'.format(tp, xs))
thisLikelihood = candidateLikelihood
if isinstance(production, Index):
theseUses = Uses(possibleVariables=possibleVariables,
actualVariables=1.,
possibleUses=possibleUses.copy(),
actualUses={})
else:
theseUses = Uses(possibleVariables=possibleVariables,
actualVariables=0.,
possibleUses=possibleUses.copy(),
actualUses={production: 1.})
# Accumulate likelihood from free variables and holes and
# arguments
for freeType, freeExpression in chain(
variableBindings.values(), zip(argumentTypes, xs)):
freeType = freeType.apply(newContext)
newContext, expressionLikelihood, newUses = self._logLikelihood(
newContext, environment, freeType, freeExpression)
if expressionLikelihood is NEGATIVEINFINITY:
thisLikelihood = NEGATIVEINFINITY
break
thisLikelihood += expressionLikelihood
theseUses += newUses
if thisLikelihood is NEGATIVEINFINITY:
continue
weightedUses.append((thisLikelihood, theseUses))
totalLikelihood = lse(totalLikelihood, thisLikelihood)
# Any of these new context objects should be equally good
context = newContext
if totalLikelihood is NEGATIVEINFINITY:
return context, totalLikelihood, Uses.empty
assert weightedUses != []
allUses = Uses.join(totalLikelihood, *weightedUses)
# memoize result
if shouldDoCaching:
outTypes = [request.apply(context)] + \
[v.apply(context) for v in environment]
outTypes = canonicalTypes(outTypes)
self.likelihoodCache[cacheKey] = (
outTypes, totalLikelihood, allUses)
return context, totalLikelihood, allUses
def expectedUses(self, frontiers):
if len(list(frontiers)) == 0:
return Uses()
likelihoods = [[(l + entry.logLikelihood, u)
for entry in frontier
for l, u in [self.closedUses(frontier.task.request, entry.program)]]
for frontier in frontiers]
zs = (lse([l for l, _ in ls]) for ls in likelihoods)
return sum(math.exp(l - z) * u
for z, frontier in zip(zs, likelihoods)
for l, u in frontier)
def insideOutside(self, frontiers, pseudoCounts):
uses = self.expectedUses(frontiers)
return FragmentGrammar(log(uses.actualVariables +
pseudoCounts) -
log(max(uses.possibleVariables, 1.)), [(log(uses.actualUses.get(p, 0.) +
pseudoCounts) -
log(uses.possibleUses.get(p, 0.) +
pseudoCounts), t, p) for _, t, p in self.productions])
def jointFrontiersLikelihood(self, frontiers):
return sum(lse([entry.logLikelihood + self.logLikelihood(frontier.task.request, entry.program)
for entry in frontier])
for frontier in frontiers)
def jointFrontiersMDL(self, frontiers, CPUs=1):
return sum(
parallelMap(
CPUs,
lambda frontier: max(
entry.logLikelihood +
self.logLikelihood(
frontier.task.request,
entry.program) for entry in frontier),
frontiers))
def __len__(self): return len(self.productions)
@staticmethod
def fromGrammar(g):
return FragmentGrammar(g.logVariable, g.productions)
def toGrammar(self):
return Grammar(self.logVariable, [(l, q.infer(), q)
for l, t, p in self.productions
for q in [defragment(p)]])
@property
def primitives(self): return [p for _, _, p in self.productions]
@staticmethod
def uniform(productions):
return FragmentGrammar(0., [(0., p.infer(), p) for p in productions])
def normalize(self):
z = lse([l for l, t, p in self.productions] + [self.logVariable])
return FragmentGrammar(self.logVariable - z,
[(l - z, t, p) for l, t, p in self.productions])
def makeUniform(self):
return FragmentGrammar(0., [(0., p.infer(), p)
for _, _, p in self.productions])
def rescoreFrontier(self, frontier):
return Frontier([FrontierEntry(e.program,
logPrior=self.logLikelihood(frontier.task.request, e.program),
logLikelihood=e.logLikelihood)
for e in frontier],
frontier.task)
@staticmethod
def induceFromFrontiers(
g0,
frontiers,
_=None,
topK=1,
topk_use_only_likelihood=False,
pseudoCounts=1.0,
aic=1.0,
structurePenalty=0.001,
a=0,
CPUs=1):
_ = topk_use_only_likelihood # not used in python compressor
originalFrontiers = frontiers
frontiers = [frontier for frontier in frontiers if not frontier.empty]
eprint("Inducing a grammar from", len(frontiers), "frontiers")
bestGrammar = FragmentGrammar.fromGrammar(g0)
oldJoint = bestGrammar.jointFrontiersMDL(frontiers, CPUs=1)
# "restricted frontiers" only contain the top K according to the best grammar
def restrictFrontiers():
return parallelMap(
CPUs,
lambda f: bestGrammar.rescoreFrontier(f).topK(topK),
frontiers)
restrictedFrontiers = []
def grammarScore(g):
g = g.makeUniform().insideOutside(restrictedFrontiers, pseudoCounts)
likelihood = g.jointFrontiersMDL(restrictedFrontiers)
structure = sum(primitiveSize(p) for p in g.primitives)
score = likelihood - aic * len(g) - structurePenalty * structure
g.clearCache()
if invalid(score):
# FIXME: This should never occur but it does anyway
score = float('-inf')
return score, g
if aic is not POSITIVEINFINITY:
restrictedFrontiers = restrictFrontiers()
bestScore, _ = grammarScore(bestGrammar)
eprint("Starting score", bestScore)
while True:
restrictedFrontiers = restrictFrontiers()
fragments = [f
for f in proposeFragmentsFromFrontiers(restrictedFrontiers, a, CPUs=CPUs)
if f not in bestGrammar.primitives
and defragment(f) not in bestGrammar.primitives]
eprint("Proposed %d fragments." % len(fragments))
candidateGrammars = [
FragmentGrammar.uniform(
bestGrammar.primitives +
[fragment]) for fragment in fragments]
if not candidateGrammars:
break
scoredFragments = parallelMap(CPUs, grammarScore, candidateGrammars,
# Each process handles up to 100
# grammars at a time, a "job"
chunksize=max(
1, min(len(candidateGrammars) // CPUs, 100)),
# maxTasks: Maximum number of jobs allocated to a process
# This means that after evaluating this*chunk many grammars,
# we killed the process, freeing up its memory.
# In exchange we pay the cost of spawning a new process.
# We should play with this number,
# figuring out how big we can make it without
# running out of memory.
maxtasksperchild=5)
newScore, newGrammar = max(scoredFragments, key=lambda sg: sg[0])
if newScore <= bestScore:
break
dS = newScore - bestScore
bestScore, bestGrammar = newScore, newGrammar
newPrimitiveLikelihood, newType, newPrimitive = bestGrammar.productions[-1]
expectedUses = bestGrammar.expectedUses(
restrictedFrontiers).actualUses.get(newPrimitive, 0)
eprint(
"New primitive of type %s\t%s\t\n(score = %f; dScore = %f; <uses> = %f)" %
(newType, newPrimitive, newScore, dS, expectedUses))
# Rewrite the frontiers in terms of the new fragment
concretePrimitive = defragment(newPrimitive)
bestGrammar.productions[-1] = (newPrimitiveLikelihood,
concretePrimitive.tp,
concretePrimitive)
frontiers = parallelMap(
CPUs, lambda frontier: bestGrammar.rescoreFrontier(
RewriteFragments.rewriteFrontier(
frontier, newPrimitive)), frontiers)
eprint(
"\t(<uses> in rewritten frontiers: %f)" %
(bestGrammar.expectedUses(frontiers).actualUses[concretePrimitive]))
else:
eprint("Skipping fragment proposals")
if False:
# Reestimate the parameters using the entire frontiers
bestGrammar = bestGrammar.makeUniform().insideOutside(frontiers, pseudoCounts)
elif True:
# Reestimate the parameters using the best programs
restrictedFrontiers = restrictFrontiers()
bestGrammar = bestGrammar.makeUniform().insideOutside(
restrictedFrontiers, pseudoCounts)
else:
# Use parameters that were found during search
pass
eprint("Old joint = %f\tNew joint = %f\n" %
(oldJoint, bestGrammar.jointFrontiersMDL(frontiers, CPUs=CPUs)))
# Return all of the frontiers, which have now been rewritten to use the
# new fragments
frontiers = {f.task: f for f in frontiers}
frontiers = [frontiers.get(f.task, f)
for f in originalFrontiers]
productionUses = bestGrammar.expectedUses(
[f for f in frontiers if not f.empty]).actualUses
productionUses = {
p: productionUses.get(
p, 0.) for p in bestGrammar.primitives}
possibleUses = bestGrammar.expectedUses(
[f for f in frontiers if not f.empty]).possibleUses
possibleUses = {
p: possibleUses.get(
p, 0.) for p in bestGrammar.primitives}
for p in bestGrammar.primitives:
eprint("%f / %f\t%s" % (productionUses[p],
possibleUses[p],
p))
bestGrammar.clearCache()
grammar = bestGrammar.toGrammar()
if False and \
any(productionUses.get(p, 0) < 0.5 for p in grammar.primitives if p.isInvented):
uselessProductions = [ p for p in grammar.primitives
if p.isInvented and productionUses.get(p, 0) < 0.5]
eprint("The following invented primitives are no longer needed, removing them...")
eprint("\t" + "\t\n".join(map(str, uselessProductions)))
grammar = grammar.removeProductions(uselessProductions)
return grammar, frontiers
|