def corrupt_example(self, e):
"""
Return a corrupted version of example e, plus the weight of this example.
"""
from hyperparameters import HYPERPARAMETERS
import random
import copy
e = copy.copy(e)
last = e[-1]
cnt = 0
while e[-1] == last:
if HYPERPARAMETERS["NGRAM_FOR_TRAINING_NOISE"] == 0:
e[-1] = random.randint(0, self.parameters.vocab_size-1)
pr = 1./self.parameters.vocab_size
elif HYPERPARAMETERS["NGRAM_FOR_TRAINING_NOISE"] == 1:
import noise
from common.myrandom import weighted_sample
e[-1], pr = weighted_sample(noise.indexed_weights())
# from vocabulary import wordmap
# print wordmap.str(e[-1]), pr
else:
assert 0
cnt += 1
# Backoff to 0gram smoothing if we fail 10 times to get noise.
if cnt > 10: e[-1] = random.randint(0, self.parameters.vocab_size-1)
weight = 1./pr
return e, weight
logging.info(stats())
# print "FINAL VALIDATION AT TRAINING STEP %d: mean(logrank) = %.2f, stddev(logrank) = %.2f, cnt = %d" % (cnt, numpy.mean(numpy.array(logranks)), numpy.std(numpy.array(logranks)), i+1)
# print stats()
if __name__ == "__main__":
import common.hyperparameters, common.options
HYPERPARAMETERS = common.hyperparameters.read("language-model")
HYPERPARAMETERS, options, args, newkeystr = common.options.reparse(HYPERPARAMETERS)
import hyperparameters
from common import myyaml
import sys
print >> sys.stderr, myyaml.dump(common.dump.vars_seq([hyperparameters, miscglobals]))
import noise
indexed_weights = noise.indexed_weights()
rundir = common.dump.create_canonical_directory(HYPERPARAMETERS)
import os.path, os
logfile = os.path.join(rundir, "log")
if newkeystr != "":
verboselogfile = os.path.join(rundir, "log%s" % newkeystr)
print >> sys.stderr, "Logging to %s, and creating link %s" % (logfile, verboselogfile)
os.system("ln -s log %s " % (verboselogfile))
else:
print >> sys.stderr, "Logging to %s, not creating any link because of default settings" % logfile
import random, numpy
random.seed(miscglobals.RANDOMSEED)
numpy.random.seed(miscglobals.RANDOMSEED)
logging.info(stats())
# print "FINAL VALIDATION AT TRAINING STEP %d: mean(logrank) = %.2f, stddev(logrank) = %.2f, cnt = %d" % (cnt, numpy.mean(numpy.array(logranks)), numpy.std(numpy.array(logranks)), i+1)
# print stats()
if __name__ == "__main__":
import common.hyperparameters, common.options
HYPERPARAMETERS = common.hyperparameters.read("language-model")
HYPERPARAMETERS, options, args, newkeystr = common.options.reparse(HYPERPARAMETERS)
import hyperparameters
from common import myyaml
import sys
print >> sys.stderr, myyaml.dump(common.dump.vars_seq([hyperparameters, miscglobals]))
import noise
indexed_weights = noise.indexed_weights()
from rundir import rundir
rundir = rundir()
import os.path, os
logfile = os.path.join(rundir, "log")
if newkeystr != "":
verboselogfile = os.path.join(rundir, "log%s" % newkeystr)
print >> sys.stderr, "Logging to %s, and creating link %s" % (logfile, verboselogfile)
os.system("ln -s log %s " % (verboselogfile))
else:
print >> sys.stderr, "Logging to %s, not creating any link because of default settings" % logfile
import random, numpy
random.seed(miscglobals.RANDOMSEED)