Fraser-Greenlee
add dataset code
3bdb76c
# analog of list.py for regex tasks. Responsible for actually running the task.
from dreamcoder.domains.regex.makeRegexTasks import makeOldTasks, makeLongTasks, makeShortTasks, makeWordTasks, makeNumberTasks, makeHandPickedTasks, makeNewTasks, makeNewNumberTasks
from dreamcoder.domains.regex.regexPrimitives import basePrimitives, altPrimitives, easyWordsPrimitives, alt2Primitives, concatPrimitives, reducedConcatPrimitives, strConstConcatPrimitives, PRC
from dreamcoder.dreamcoder import explorationCompression, Task
from dreamcoder.grammar import Grammar
from dreamcoder.likelihoodModel import add_cutoff_values, add_string_constants
from dreamcoder.program import Abstraction, Application
from dreamcoder.type import tpregex
from dreamcoder.utilities import eprint, flatten, testTrainSplit, POSITIVEINFINITY
import random
import math
import pregex as pre
import os
try:
from dreamcoder.recognition import RecurrentFeatureExtractor, JSONFeatureExtractor
class LearnedFeatureExtractor(RecurrentFeatureExtractor):
H = 64
special = 'regex'
def tokenize(self, examples):
def sanitize(l): return [z if z in self.lexicon else "?"
for z_ in l
for z in (z_ if isinstance(z_, list) else [z_])]
tokenized = []
for xs, y in examples:
if isinstance(y, list):
y = ["LIST_START"] + y + ["LIST_END"]
else:
y = [y]
y = sanitize(y)
if len(y) > self.maximumLength:
return None
serializedInputs = []
for xi, x in enumerate(xs):
if isinstance(x, list):
x = ["LIST_START"] + x + ["LIST_END"]
else:
x = [x]
x = sanitize(x)
if len(x) > self.maximumLength:
return None
serializedInputs.append(x)
tokenized.append((tuple(serializedInputs), y))
return tokenized
def __init__(self, tasks, testingTasks=[], cuda=False):
self.lexicon = set(flatten((t.examples for t in tasks + testingTasks), abort=lambda x: isinstance(
x, str))).union({"LIST_START", "LIST_END", "?"})
self.num_examples_list = [len(t.examples) for t in tasks]
# Calculate the maximum length
self.maximumLength = POSITIVEINFINITY
self.maximumLength = max(len(l)
for t in tasks + testingTasks
for xs, y in self.tokenize(t.examples)
for l in [y] + [x for x in xs])
super(
LearnedFeatureExtractor,
self).__init__(
lexicon=list(
self.lexicon),
tasks=tasks,
cuda=cuda,
H=self.H,
bidirectional=True)
self.parallelTaskOfProgram = False
def taskOfProgram(self, p, t):
#raise NotImplementedError
num_examples = random.choice(self.num_examples_list)
p = p.visit(ConstantInstantiateVisitor.SINGLE)
preg = p.evaluate([])(pre.String(""))
t = Task("Helm", t, [((), list(preg.sample())) for _ in range(num_examples) ])
return t
except: pass
#in init: loop over tasks, save lengths,
class ConstantInstantiateVisitor(object):
def __init__(self):
self.regexes = [
pre.create(".+"),
pre.create("\d+"),
pre.create("\w+"),
pre.create("\s+"),
pre.create("\\u+"),
pre.create("\l+")]
def primitive(self, e):
if e.name == "r_const":
#return Primitive("STRING", e.tp, random.choice(self.words))
s = random.choice(self.regexes).sample() #random string const
s = pre.String(s)
e.value = PRC(s,arity=0)
return e
def invented(self, e): return e.body.visit(self)
def index(self, e): return e
def application(self, e):
return Application(e.f.visit(self), e.x.visit(self))
def abstraction(self, e):
return Abstraction(e.body.visit(self))
#TODO fix
class MyJSONFeatureExtractor(JSONFeatureExtractor):
N_EXAMPLES = 5
def _featuresOfProgram(self, program, tp):
try:
preg = program.evaluate([])
# if 'left_paren' in program.show(False):
#eprint("string_pregex:", string_pregex)
#eprint("string_pregex:", string_pregex)
except IndexError:
# free variable
return None
except Exception as e:
eprint("Exception during evaluation:", e)
if "Attempt to evaluate fragment variable" in e:
eprint("program (bc fragment error)", program)
return None
examples = []
for _ in range(self.N_EXAMPLES * 5): # oh this is arbitrary ig
try:
y = preg.sample() # TODO
#this line should keep inputs short, so that helmholtzbatch can be large
#allows it to try other samples
#(Could also return None off the bat... idk which is better)
#if len(y) > 20:
# continue
#eprint(tp, program, x, y)
examples.append(y)
except BaseException:
continues
if len(examples) >= self.N_EXAMPLES:
break
else:
return None
return examples # changed to list_features(examples) from examples
def regex_options(parser):
parser.add_argument("--maxTasks", type=int,
default=500,
help="truncate tasks to fit within this boundary")
parser.add_argument(
"--maxExamples",
type=int,
default=10,
help="truncate number of examples per task to fit within this boundary")
parser.add_argument("--tasks",
default="long",
help="which tasks to use",
choices=["old", "short", "long", "words", "number", "handpicked", "new", "newNumber"])
parser.add_argument("--primitives",
default="concat",
help="Which primitive set to use",
choices=["base", "alt1", "easyWords", "alt2", "concat", "reduced", "strConst"])
parser.add_argument("--extractor", type=str,
choices=["hand", "deep", "learned", "json"],
default="learned") # if i switch to json it breaks
parser.add_argument("--split", metavar="TRAIN_RATIO",
type=float,
default=0.8,
help="split test/train")
parser.add_argument("-H", "--hidden", type=int,
default=256,
help="number of hidden units")
parser.add_argument("--likelihoodModel",
default="probabilistic",
help="likelihood Model",
choices=["probabilistic", "all-or-nothing"])
parser.add_argument("--topk_use_map",
dest="topk_use_only_likelihood",
action="store_false")
parser.add_argument("--debug",
dest="debug",
action="store_true")
parser.add_argument("--ll_cutoff",
dest="use_ll_cutoff",
nargs='*',
default=False,
help="use ll cutoff for training tasks (for probabilistic likelihood model only). default is False,")
parser.add_argument("--use_str_const",
action="store_true",
help="use string constants")
"""parser.add_argument("--stardecay",
type=float,
dest="stardecay",
default=0.5,
help="p value for kleenestar and plus")"""
# Lucas recommends putting a struct with the definitions of the primitives here.
# TODO:
# Build likelihood funciton
# modify NN
# make primitives
# make tasks
def main(args):
"""
Takes the return value of the `commandlineArguments()` function as input and
trains/tests the model on regular expressions.
"""
#for dreaming
#parse use_ll_cutoff
use_ll_cutoff = args.pop('use_ll_cutoff')
if not use_ll_cutoff is False:
#if use_ll_cutoff is a list of strings, then train_ll_cutoff and train_ll_cutoff
#will be tuples of that string followed by the actual model
if len(use_ll_cutoff) == 1:
train_ll_cutoff = use_ll_cutoff[0] # make_cutoff_model(use_ll_cutoff[0], tasks))
test_ll_cutoff = use_ll_cutoff[0] # make_cutoff_model(use_ll_cutoff[0], tasks))
else:
assert len(use_ll_cutoff) == 2
train_ll_cutoff = use_ll_cutoff[0] #make_cutoff_model(use_ll_cutoff[0], tasks))
test_ll_cutoff = use_ll_cutoff[1] #make_cutoff_model(use_ll_cutoff[1], tasks))
else:
train_ll_cutoff = None
test_ll_cutoff = None
regexTasks = {"old": makeOldTasks,
"short": makeShortTasks,
"long": makeLongTasks,
"words": makeWordTasks,
"number": makeNumberTasks,
"handpicked": makeHandPickedTasks,
"new": makeNewTasks,
"newNumber": makeNewNumberTasks
}[args.pop("tasks")]
tasks = regexTasks() # TODO
eprint("Generated", len(tasks), "tasks")
maxTasks = args.pop("maxTasks")
if len(tasks) > maxTasks:
eprint("Unwilling to handle {} tasks, truncating..".format(len(tasks)))
seed = 42 # previously this was hardcoded and never changed
random.seed(seed)
random.shuffle(tasks)
del tasks[maxTasks:]
maxExamples = args.pop("maxExamples")
split = args.pop("split")
test, train = testTrainSplit(tasks, split)
eprint("Split tasks into %d/%d test/train" % (len(test), len(train)))
test = add_cutoff_values(test, test_ll_cutoff)
train = add_cutoff_values(train, train_ll_cutoff)
eprint("added cutoff values to tasks, train: ", train_ll_cutoff, ", test:", test_ll_cutoff )
if args.pop("use_str_const"):
assert args["primitives"] == "strConst" or args["primitives"] == "reduced"
ConstantInstantiateVisitor.SINGLE = \
ConstantInstantiateVisitor()
test = add_string_constants(test)
train = add_string_constants(train)
eprint("added string constants to test and train")
for task in test + train:
if len(task.examples) > maxExamples:
task.examples = task.examples[:maxExamples]
task.specialTask = ("regex", {"cutoff": task.ll_cutoff, "str_const": task.str_const})
task.examples = [(xs, [y for y in ys ])
for xs,ys in task.examples ]
task.maxParameters = 1
# from list stuff
primtype = args.pop("primitives")
prims = {"base": basePrimitives,
"alt1": altPrimitives,
"alt2": alt2Primitives,
"easyWords": easyWordsPrimitives,
"concat": concatPrimitives,
"reduced": reducedConcatPrimitives,
"strConst": strConstConcatPrimitives
}[primtype]
extractor = {
"learned": LearnedFeatureExtractor,
"json": MyJSONFeatureExtractor
}[args.pop("extractor")]
extractor.H = args.pop("hidden")
#stardecay = args.stardecay
#stardecay = args.pop('stardecay')
#decaystr = 'd' + str(stardecay)
import datetime
timestamp = datetime.datetime.now().isoformat()
outputDirectory = "experimentOutputs/regex/%s"%timestamp
os.system("mkdir -p %s"%outputDirectory)
args.update({
"featureExtractor": extractor,
"outputPrefix": "%s/regex"%(outputDirectory),
"evaluationTimeout": 0.005,
"topk_use_only_likelihood": True,
"maximumFrontier": 10,
"compressor": args.get("compressor","ocaml")
})
####
# use the
#prim_list = prims(stardecay)
prim_list = prims()
specials = ["r_kleene", "r_plus", "r_maybe", "r_alt", "r_concat"]
n_base_prim = len(prim_list) - len(specials)
productions = [
(math.log(0.5 / float(n_base_prim)),
prim) if prim.name not in specials else (
math.log(0.10),
prim) for prim in prim_list]
baseGrammar = Grammar.fromProductions(productions, continuationType=tpregex)
#baseGrammar = Grammar.uniform(prims())
#for i in range(100):
# eprint(baseGrammar.sample(tpregex))
#eprint(baseGrammar)
#explore
test_stuff = args.pop("debug")
if test_stuff:
eprint(baseGrammar)
eprint("sampled programs from prior:")
for i in range(100): #100
eprint(baseGrammar.sample(test[0].request,maximumDepth=1000))
eprint("""half the probability mass is on higher-order primitives.
Therefore half of enumerated programs should have more than one node.
However, we do not observe this.
Instead we see a very small fraction of programs have more than one node.
So something seems to be wrong with grammar.sample.
Furthermore: observe the large print statement above.
This prints the candidates for sampleDistribution in grammar.sample.
the first element of each tuple is the probability passed into sampleDistribution.
Half of the probability mass should be on the functions, but instead they are equally
weighted with the constants. If you look at the grammar above, this is an error!!!!
""")
assert False
del args["likelihoodModel"]
explorationCompression(baseGrammar, train,
testingTasks = test,
**args)