forked from ellisk42/ec
-
Notifications
You must be signed in to change notification settings - Fork 0
/
ec.py
549 lines (499 loc) · 26.7 KB
/
ec.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
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
from utilities import eprint
from likelihoodModel import *
from recognition import *
from frontier import *
from program import *
from type import *
from task import *
from enumeration import *
from grammar import *
from fragmentGrammar import *
import baselines
import dill
import os
import datetime
import cPickle as pickle
import torch
class ECResult():
def __init__(self, _=None,
learningCurve=None,
grammars=None,
taskSolutions=None,
averageDescriptionLength=None,
parameters=None,
recognitionModel=None,
searchTimes=None,
baselines=None):
self.searchTimes = searchTimes or []
self.recognitionModel = recognitionModel
self.averageDescriptionLength = averageDescriptionLength or []
self.parameters = parameters
self.learningCurve = learningCurve or []
self.grammars = grammars or []
self.taskSolutions = taskSolutions or {}
# baselines is a dictionary of name -> ECResult
self.baselines = baselines or {}
def __repr__(self):
attrs = ["{}={}".format(k, v) for k, v in self.__dict__.items()]
return "ECResult({})".format(", ".join(attrs))
# Linux does not like files that have more than 256 characters
# So when exporting the results we abbreviate the parameters
abbreviations = {"frontierSize": "fs",
"iterations": "it",
"maximumFrontier": "MF",
"onlyBaselines": "baseline",
"pseudoCounts": "pc",
"structurePenalty": "L",
"helmholtzRatio": "HR",
"topK": "K",
"enumerationTimeout": "ET",
"useRecognitionModel": "rec"}
@staticmethod
def abbreviate(parameter): return ECResult.abbreviations.get(parameter, parameter)
@staticmethod
def parameterOfAbbreviation(abbreviation):
return ECResult.abbreviationToParameter.get(abbreviation, abbreviation)
ECResult.abbreviationToParameter = {v:k for k,v in ECResult.abbreviations.iteritems() }
def explorationCompression(*arguments, **keywords):
for r in ecIterator(*arguments, **keywords): pass
return r
def ecIterator(grammar, tasks,
_=None,
bootstrap=None,
solver="ocaml",
compressor="rust",
likelihoodModel="all-or-nothing",
testingTasks = [],
benchmark=None,
iterations=None,
resume=None,
frontierSize=None,
enumerationTimeout=None,
expandFrontier=None,
resumeFrontierSize=None,
useRecognitionModel=False,
useNewRecognitionModel=True,
steps=250,
helmholtzRatio=0.,
helmholtzBatch=5000,
featureExtractor = None,
activation='relu',
topK=1,
maximumFrontier=None,
pseudoCounts=1.0, aic=1.0,
structurePenalty=0.001, arity=0,
evaluationTimeout=0.05, # seconds
CPUs=1,
cuda=False,
message="",
onlyBaselines=False,
outputPrefix=None):
if frontierSize is None and enumerationTimeout is None:
eprint("Please specify a frontier size and/or an enumeration timeout:",
"explorationCompression(..., enumerationTimeout = ..., frontierSize = ...)")
assert False
if iterations is None:
eprint("Please specify a iteration count: explorationCompression(..., iterations = ...)")
assert False
if useRecognitionModel and featureExtractor is None:
eprint("Warning: Recognition model needs feature extractor.",
"Ignoring recognition model.")
useRecognitionModel = False
if useNewRecognitionModel and featureExtractor is None:
eprint("Warning: Recognition model needs feature extractor.",
"Ignoring recognition model.")
useNewRecognitionModel = False
if benchmark is not None and resume is None:
eprint("You cannot benchmark unless you are loading a checkpoint, aborting.")
assert False
# We save the parameters that were passed into EC
# This is for the purpose of exporting the results of the experiment
parameters = {k: v for k, v in locals().iteritems()
if k not in {"tasks", "grammar", "cuda", "_", "solver",
"message", "CPUs", "outputPrefix",
"resume", "resumeFrontierSize", "bootstrap",
"featureExtractor", "benchmark",
"evaluationTimeout", "testingTasks", "compressor"} \
and v is not None}
if not useRecognitionModel:
for k in {"activation","helmholtzRatio","steps"}: del parameters[k]
# Uses `parameters` to construct the checkpoint path
def checkpointPath(iteration, extra=""):
parameters["iterations"] = iteration
kvs = ["{}={}".format(ECResult.abbreviate(k), parameters[k]) for k in sorted(parameters.keys())]
if useRecognitionModel or useNewRecognitionModel:
kvs += ["feat=%s"%(featureExtractor.__name__)]
if bootstrap:
kvs += ["bstrap=True"]
return "{}_{}{}.pickle".format(outputPrefix, "_".join(kvs), extra)
if onlyBaselines and not benchmark:
result = ECResult()
result.baselines = baselines.all(grammar, tasks,
CPUs=CPUs, cuda=cuda, featureExtractor=featureExtractor,
**parameters)
if outputPrefix is not None:
path = checkpointPath(0, extra="_baselines")
with open(path, "wb") as f:
pickle.dump(result, f)
eprint("Exported checkpoint to", path)
yield result
return
if message: message = " ("+message+")"
eprint("Running EC%s on %s @ %s with %d CPUs and parameters:"%(message, os.uname()[1],
datetime.datetime.now(), CPUs))
for k,v in parameters.iteritems():
eprint("\t", k, " = ", v)
eprint("\t", "evaluationTimeout", " = ", evaluationTimeout)
eprint()
# Restore checkpoint
if resume is not None:
path = checkpointPath(resume, extra = "_baselines" if onlyBaselines else "")
with open(path, "rb") as handle:
result = pickle.load(handle)
eprint("Loaded checkpoint from", path)
grammar = result.grammars[-1] if result.grammars else grammar
recognizer = result.recognitionModel
if resumeFrontierSize:
frontierSize = resumeFrontierSize
eprint("Set frontier size to", frontierSize)
if bootstrap is not None: # Make sure that we register bootstrapped primitives
for p in grammar.primitives: RegisterPrimitives.register(p)
else: # Start from scratch
if bootstrap is not None:
with open(bootstrap, "rb") as handle: strapping = pickle.load(handle).grammars[-1]
eprint("Bootstrapping from",bootstrap)
eprint("Bootstrap primitives:")
for p in strapping.primitives:
eprint(p)
RegisterPrimitives.register(p)
eprint()
grammar = Grammar.uniform(list({p for p in grammar.primitives + strapping.primitives
if not str(p).startswith("fix")}))
if compressor == "rust":
eprint("Rust compressor is currently not compatible with bootstrapping.",
"Falling back on pypy compressor.")
compressor = "pypy"
result = ECResult(parameters=parameters, grammars=[grammar],
taskSolutions = { t: Frontier([], task = t) for t in tasks },
recognitionModel = None)
#just plopped this in here, hope it works: -it doesn't. having issues.
if useNewRecognitionModel and (not hasattr(result, 'recognitionModel') or type(result.recognitionModel) is not NewRecognitionModel):
eprint("Creating new recognition model")
featureExtractorObject = featureExtractor(tasks + testingTasks)
result.recognitionModel = NewRecognitionModel(featureExtractorObject, grammar, cuda=cuda)
#end
if benchmark is not None:
assert resume is not None, "Benchmarking requires resuming from checkpoint that you are benchmarking."
if benchmark > 0:
assert testingTasks != [], "Benchmarking requires held out test tasks"
benchmarkTasks = testingTasks
else:
benchmarkTasks = tasks
benchmark = -benchmark
if len(result.baselines) == 0: results = {"our algorithm": result}
else: results = result.baselines
for name, result in results.iteritems():
eprint("Starting benchmark:",name)
benchmarkSynthesisTimes(result, benchmarkTasks, timeout = benchmark, CPUs = CPUs)
eprint("Completed benchmark.")
eprint()
yield None
return
#may need to change this if it doesn't do what I need
likelihoodModel = {
"all-or-nothing": lambda: AllOrNothingLikelihoodModel(
timeout=evaluationTimeout),
"feature-discriminator": lambda: FeatureDiscriminatorLikelihoodModel(
tasks, featureExtractor(tasks)),
"euclidean": lambda: EuclideanLikelihoodModel(
featureExtractor(tasks)),
"probabilistic": lambda: ProbabilisticLikelihoodModel(timeout=evaluationTimeout) #TODO
}[likelihoodModel]()
for j in range(resume or 0, iterations):
if j >= 2 and expandFrontier and result.learningCurve[-1] <= result.learningCurve[-2]:
oldEnumerationTimeout = enumerationTimeout
if expandFrontier <= 10:
enumerationTimeout = int(enumerationTimeout * expandFrontier)
else:
enumerationTimeout = int(enumerationTimeout + expandFrontier)
eprint("Expanding enumeration timeout from {} to {} because of no progress".format(
oldEnumerationTimeout, enumerationTimeout))
#the line below may be an issue... enumeration doesn't use the recognition model
frontiers, times = multithreadedEnumeration(grammar, tasks, likelihoodModel,
solver=solver,
frontierSize=frontierSize,
maximumFrontier=maximumFrontier,
enumerationTimeout=enumerationTimeout,
CPUs=CPUs,
evaluationTimeout=evaluationTimeout)
if expandFrontier and j > 0 and (not useRecognitionModel) and \
sum(not f.empty for f in frontiers) <= result.learningCurve[-1]:
timeout = enumerationTimeout
unsolvedTasks = [ f.task for f in frontiers if f.empty ]
while True:
eprint("Expanding enumeration timeout from %i to %i because of no progress. Focusing exclusively on %d unsolved tasks."%(
timeout, timeout*expandFrontier, len(unsolvedTasks)))
timeout = timeout*expandFrontier
unsolvedFrontiers, unsolvedTimes = \
multithreadedEnumeration(grammar, unsolvedTasks, likelihoodModel,
solver=solver,
frontierSize=frontierSize,
maximumFrontier=maximumFrontier,
enumerationTimeout=timeout,
CPUs=CPUs,
evaluationTimeout=evaluationTimeout)
if any( not f.empty for f in unsolvedFrontiers ):
times += unsolvedTimes
unsolvedFrontiers = {f.task: f for f in unsolvedFrontiers }
frontiers = [ f if not f.empty else unsolvedFrontiers[f.task]
for f in frontiers ]
break
eprint("Generative model enumeration results:")
eprint(Frontier.describe(frontiers))
tasksHitTopDown = {f.task for f in frontiers if not f.empty}
# Train + use recognition model
if useRecognitionModel:
featureExtractorObject = featureExtractor(tasks)
recognizer = RecognitionModel(featureExtractorObject, grammar, activation=activation, cuda=cuda)
##transplanted test code
#requests = [ frontier.task.request for frontier in frontiers ]
#eprint("requests:")
#eprint(requests)
recognizer.train(frontiers, topK=topK, steps=steps,
CPUs=CPUs,
helmholtzBatch=helmholtzBatch,
helmholtzRatio=helmholtzRatio if j > 0 else 0.)
result.recognitionModel = recognizer
bottomupFrontiers, times = result.recognitionModel.enumerateFrontiers(tasks, likelihoodModel,
CPUs=CPUs,
solver=solver,
maximumFrontier=maximumFrontier,
frontierSize=frontierSize,
enumerationTimeout=enumerationTimeout,
evaluationTimeout=evaluationTimeout)
elif useNewRecognitionModel: # Train a recognition model
result.recognitionModel.updateGrammar(grammar)
result.recognitionModel.train(frontiers, topK=topK, steps=steps, helmholtzRatio=helmholtzRatio) #changed from result.frontiers to frontiers and added thingy
eprint("done training recognition model")
bottomupFrontiers = result.recognitionModel.enumerateFrontiers(tasks, likelihoodModel,
CPUs=CPUs,
solver=solver,
maximumFrontier=maximumFrontier,
frontierSize=frontierSize,
enumerationTimeout=enumerationTimeout,
evaluationTimeout=evaluationTimeout)
if useRecognitionModel or useNewRecognitionModel:
eprint("Recognition model enumeration results:")
eprint(Frontier.describe(bottomupFrontiers))
result.averageDescriptionLength.append(mean( -f.marginalLikelihood()
for f in bottomupFrontiers
if not f.empty ))
tasksHitBottomUp = {f.task for f in bottomupFrontiers if not f.empty}
showHitMatrix(tasksHitTopDown, tasksHitBottomUp, tasks)
# Rescore the frontiers according to the generative model
# and then combine w/ original frontiers
bottomupFrontiers = [ grammar.rescoreFrontier(f) for f in bottomupFrontiers ]
frontiers = [f.combine(b) for f, b in zip(frontiers, bottomupFrontiers)]
else:
result.averageDescriptionLength.append(mean( -f.marginalLikelihood()
for f in frontiers
if not f.empty ))
if useRecognitionModel:
result.searchTimes.append(times)
eprint("Average search time: ",int(mean(times)+0.5),
"sec.\tmedian:",int(median(times)+0.5),
"\tmax:",int(max(times)+0.5),
"\tstandard deviation",int(standardDeviation(times)+0.5))
# Incorporate frontiers from anything that was not hit
frontiers = [ f if not f.empty
else grammar.rescoreFrontier(result.taskSolutions.get(f.task, Frontier.makeEmpty(f.task)))
for f in frontiers ]
frontiers = [ f.topK(maximumFrontier) for f in frontiers ]
if maximumFrontier <= 10:
eprint("Because maximumFrontier is small (<=10), I am going to show you the full contents of all the frontiers:")
for f in frontiers:
if f.empty: continue
eprint(f.task)
for e in f.normalize():
eprint("%.02f\t%s"%(e.logPosterior, e.program))
eprint()
# Record the new solutions
result.taskSolutions = {f.task: f.topK(topK)
for f in frontiers}
result.learningCurve += [sum(f is not None and not f.empty for f in result.taskSolutions.values() )]
# Sleep-G
#eprint("frontiers:")
#eprint(frontiers)
grammar, frontiers = induceGrammar(grammar, frontiers,
topK=topK, pseudoCounts=pseudoCounts, a=arity,
aic=aic, structurePenalty=structurePenalty,
backend=compressor, CPUs=CPUs)
result.grammars.append(grammar)
eprint("Grammar after iteration %d:" % (j + 1))
eprint(grammar)
eprint("Expected uses of each grammar production after iteration %d:" % (j + 1))
productionUses = FragmentGrammar.fromGrammar(grammar).\
expectedUses([f for f in frontiers if not f.empty ]).actualUses
productionUses = {p: productionUses.get(p,0.) for p in grammar.primitives }
for p in sorted(productionUses.keys(),key = lambda p: -productionUses[p]):
eprint("<uses>=%.2f\t%s"%(productionUses[p], p))
eprint()
if maximumFrontier <= 10:
eprint("Because maximumFrontier is small (<=10), I am going to show you the full contents of all the rewritten frontiers:")
for f in frontiers:
if f.empty: continue
eprint(f.task)
for e in f.normalize():
eprint("%.02f\t%s"%(e.logPosterior, e.program))
eprint()
if outputPrefix is not None:
path = checkpointPath(j + 1)
with open(path, "wb") as handle:
try:
dill.dump(result, handle)
except TypeError as e:
eprint(result)
assert(False)
eprint("Exported checkpoint to", path)
yield result
def showHitMatrix(top, bottom, tasks):
tasks = set(tasks)
total = bottom|top
eprint(len(total),"/",len(tasks),"total hit tasks")
bottomMiss = tasks - bottom
topMiss = tasks - top
eprint("{: <13s}{: ^13s}{: ^13s}".format("","bottom miss","bottom hit"))
eprint("{: <13s}{: ^13d}{: ^13d}".format("top miss",
len(bottomMiss & topMiss),
len(bottom & topMiss)))
eprint("{: <13s}{: ^13d}{: ^13d}".format("top hit",
len(top & bottomMiss),
len(top & bottom)))
def commandlineArguments(_=None,
iterations=None,
frontierSize=None,
enumerationTimeout=None,
topK=1,
CPUs=1,
useRecognitionModel=False,
useNewRecognitionModel=True,
steps=2,
activation='relu',
helmholtzRatio=0.,
helmholtzBatch=5000,
featureExtractor = None,
cuda=None,
maximumFrontier=None,
pseudoCounts=1.0, aic=1.0,
structurePenalty=0.001, a=0,
onlyBaselines=False,
extras=None):
if cuda is None:
cuda = torch.cuda.is_available()
import argparse
parser = argparse.ArgumentParser(description="")
parser.add_argument("--resume",
help="Resumes EC algorithm from checkpoint",
default=None,
type=int)
parser.add_argument("-i", "--iterations",
help="default: %d" % iterations,
default=iterations,
type=int)
parser.add_argument("-f", "--frontierSize",
default=frontierSize,
help="default: %s" % frontierSize,
type=int)
parser.add_argument("-t", "--enumerationTimeout",
default=enumerationTimeout,
help="In seconds. default: %s" % enumerationTimeout,
type=int)
parser.add_argument("-F", "--expandFrontier", metavar="FACTOR-OR-AMOUNT",
default=None,
help="if an iteration passes where no new tasks have been solved, the frontier is expanded. If the given value is less than 10, it is scaled (e.g. 1.5), otherwise it is grown (e.g. 2000).",
type=float)
parser.add_argument("--resumeFrontierSize", type=int,
help="when resuming a checkpoint which expanded the frontier, use this option to set the appropriate frontier size for the next iteration.")
parser.add_argument("-k", "--topK",
default=topK,
help="When training generative and discriminative models, we train them to fit the top K programs. Ideally we would train them to fit the entire frontier, but this is often intractable. default: %d" % topK,
type=int)
parser.add_argument("-p", "--pseudoCounts",
default=pseudoCounts,
help="default: %f" % pseudoCounts,
type=float)
parser.add_argument("-b", "--aic",
default=aic,
help="default: %f" % aic,
type=float)
parser.add_argument("-l", "--structurePenalty",
default=structurePenalty,
help="default: %f" % structurePenalty,
type=float)
parser.add_argument("-a", "--arity",
default=a,
help="default: %d" % a,
type=int)
parser.add_argument("-c", "--CPUs",
default=CPUs,
help="default: %d" % CPUs,
type=int)
parser.add_argument("--no-cuda",
action="store_false",
dest="cuda",
help="""cuda will be used if available (which it %s),
unless this is set""" % ("IS" if cuda else "ISN'T"))
parser.add_argument("-m", "--maximumFrontier",
help="""Even though we enumerate --frontierSize
programs, we might want to only keep around the very
best for performance reasons. This is a cut off on the
maximum size of the frontier that is kept around.
Default: %s""" % maximumFrontier,
type=int)
parser.add_argument("--benchmark",
help = """Benchmark synthesis times with a timeout of this many seconds. You must use the --resume option. EC will not run but instead we were just benchmarked the synthesis times of a learned model""",
type=float,
default = None)
parser.add_argument("--recognition",
dest="useRecognitionModel",
action="store_true",
help="""Enable bottom-up neural recognition model.
Default: %s""" % useRecognitionModel)
parser.add_argument("-g", "--no-recognition",
dest="useRecognitionModel",
action="store_false",
help="""Disable bottom-up neural recognition model.
Default: %s""" % (not useRecognitionModel))
parser.add_argument("--steps", type=int,
default=steps,
help="""Trainings steps for neural recognition model.
Default: %s""" % steps)
parser.add_argument("--activation",
choices=["relu", "sigmoid", "tanh"],
default=activation,
help="""Activation function for neural recognition model.
Default: %s""" % activation)
parser.add_argument("-r","--Helmholtz",
dest="helmholtzRatio",
help="""When training recognition models, what fraction of the training data should be samples from the generative model? Default %f""" % helmholtzRatio,
default = helmholtzRatio,
type=float)
parser.add_argument("--helmholtzBatch",
dest="helmholtzBatch",
help="""When training recognition models, size of the Helmholtz batch? Default %f""" % helmholtzBatch,
default=helmholtzBatch,
type=float)
parser.add_argument("-B", "--baselines", dest="onlyBaselines", action="store_true",
help="only compute baselines")
parser.add_argument("--bootstrap",
help="Start the learner out with a pretrained DSL. This argument should be a path to a checkpoint file.",
default=None,
type=str)
parser.set_defaults(useRecognitionModel=useRecognitionModel,
featureExtractor=featureExtractor,
maximumFrontier=maximumFrontier,
cuda=cuda)
if extras is not None:
extras(parser)
v = vars(parser.parse_args())
return v