def run(*args):
#print "# Running data"
global hypotheses
data_size = args[0]
p_representation = defaultdict(int) # how often do you get the right representation
p_response = defaultdict(int) # how often do you get the right response?
p_representation_literal = defaultdict(int) # how often do you get the right representation
p_response_literal = defaultdict(int) # how often do you get the right response?
p_representation_presup = defaultdict(int) # how often do you get the right representation
p_response_presup = defaultdict(int) # how often do you get the right response?
#print "# Generating data"
data = generate_data(data_size)
# recompute these
#print "# Computing posterior"
#[ x.unclear_functions() for x in hypotheses ]
[ x.compute_posterior(data) for x in hypotheses ]
# normalize the posterior in fs
#print "# Computing normalizer"
Z = logsumexp([x.posterior_score for x in hypotheses])
# and output the top hypotheses
qq = FiniteBestSet(max=True, N=25)
for h in hypotheses: qq.push(h, h.posterior_score) # get the tops
for i, h in enumerate(qq.get_all(sorted=True)):
for w in h.all_words():
fprintn(8, data_size, i, w, h.posterior_score, q(h.value[w]), f=options.OUT_PATH+"-hypotheses."+str(get_rank())+".txt")
# and compute the probability of being correct
#print "# Computing correct probability"
for h in hypotheses:
hstr = str(h)
#print data_size, len(data), exp(h.posterior_score), correct[ str(h)+":"+w ]
for w in words:
p = exp(h.posterior_score - Z)
key = w + ":" + hstr
p_representation[w] += p * (agree_pct[key] == 1.)
p_representation_presup[w] += p * (agree_pct_presup[key] == 1.) # if we always agree with the target, then we count as the right rep.
p_representation_literal[w] += p * (agree_pct_literal[key] == 1.)
# and just how often does the hypothesis agree?
p_response[w] += p * agree_pct[key]
p_response_presup[w] += p * agree_pct_presup[key]
p_response_literal[w] += p * agree_pct_literal[key]
#print "# Outputting"
for w in words:
fprintn(10, str(get_rank()), q(w), data_size, p_representation[w], p_representation_presup[w], p_representation_literal[w], p_response[w], p_response_presup[w], p_response_literal[w], f=options.OUT_PATH+"-stats."+str(get_rank())+".txt")
return 0
def run(llt=1.0):
h0 = CCGLexicon(make_hypothesis, words=all_words, alpha=0.9, palpha=0.9, likelihood_temperature=llt)
fbs = FiniteBestSet(N=10)
from LOTlib.Inference.MetropolisHastings import mh_sample
for h in lot_iter(mh_sample(h0, data, SAMPLES)):
fbs.add(h, h.posterior_score)
return fbs
def run(llt=1.0):
h0 = CCGLexicon(make_hypothesis, words=all_words, alpha=0.9, palpha=0.9, likelihood_temperature=llt)
fbs = FiniteBestSet(N=10)
from LOTlib.Inference.MetropolisHastings import mh_sample
for h in lot_iter(mh_sample(h0, data, SAMPLES)):
fbs.add(h, h.posterior_score)
return fbs
def run(*args):
"""The running function."""
# starting hypothesis -- here this generates at random
h0 = GaussianLOTHypothesis(grammar)
# We store the top 100 from each run
pq = FiniteBestSet(N=100, max=True, key="posterior_score")
pq.add(MHSampler(h0, data, STEPS, skip=SKIP))
return pq
def run(*args):
# starting hypothesis -- here this generates at random
h0 = GaussianLOTHypothesis(grammar, prior_temperature=PRIOR_TEMPERATURE)
# We store the top 100 from each run
pq = FiniteBestSet(100, max=True, key="posterior_score")
pq.add( mh_sample(h0, data, STEPS, skip=SKIP) )
return pq
def run(*args):
# starting hypothesis -- here this generates at random
h0 = GaussianLOTHypothesis(grammar,
prior_temperature=PRIOR_TEMPERATURE)
# We store the top 100 from each run
pq = FiniteBestSet(100, max=True, key="posterior_score")
pq.add(mh_sample(h0, data, STEPS, skip=SKIP))
return pq
def load_finite_trees(f): """ Load in either a list of fintie trees This can come in either two formats -- a "finite sample" of hypotheses, or a "finite sample" of lexica. In the latter case, we take the top for *each word* and just use them """ inh = open(f) fs = pickle.load(inh) if isinstance(fs.Q[0], VectorizedLexicon) or isinstance(fs.Q[0], SimpleLexicon): # else we want to extract the treesed from this priority queue upq = FiniteBestSet() for l in fs.get_all(): for e in l.dexpr.values(): upq.push(e, 0.0) return upq.get_all() else: #print type(type(fs.Q[0]).__name__) return fs.get_all()
def run(data_size):
print "Running ", data_size
# We store the top 100 from each run
hypset = FiniteBestSet(TOP_COUNT, max=True)
# initialize the data
data = generate_data(data_size)
# starting hypothesis -- here this generates at random
learner = GriceanQuantifierLexicon(make_my_hypothesis, my_weight_function)
# We will defautly generate from null the grammar if no value is specified
for w in target.all_words(): learner.set_word(w)
# populate the finite sample by running the sampler for this many steps
for x in mh_sample(learner, data, SAMPLES, skip=0):
hypset.push(x, x.posterior_score)
return hypset
def run(data_size):
print "Running ", data_size
# We store the top 100 from each run
hypset = FiniteBestSet(TOP_COUNT, max=True)
# initialize the data
data = generate_data(data_size)
# starting hypothesis -- here this generates at random
learner = GriceanQuantifierLexicon(make_my_hypothesis, my_weight_function)
# We will defautly generate from null the grammar if no value is specified
for w in target.all_words():
learner.set_word(w)
# populate the finite sample by running the sampler for this many steps
for x in mh_sample(learner, data, SAMPLES, skip=0):
hypset.push(x, x.posterior_score)
return hypset
learner = GriceanQuantifierLexicon(make_my_hypothesis, my_weight_function)
# We will defautly generate from null the grammar if no value is specified
for w in target.all_words():
learner.set_word(w)
# populate the finite sample by running the sampler for this many steps
for x in mh_sample(learner, data, SAMPLES, skip=0):
hypset.push(x, x.posterior_score)
return hypset
if __name__ == "__main__":
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# MPI interface
# Map. SimpleMPI will use a normal MAP if we are not running in MPI
allret = MPI_map(run, map(lambda x: [x],
DATA_AMOUNTS * CHAINS)) # this many chains
## combine into a single hypothesis set and save
outhyp = FiniteBestSet(max=True)
for r in allret:
print "# Merging ", len(r)
outhyp.merge(r)
import pickle
pickle.dump(outhyp, open(OUT_PATH, 'w'))
def run(*args):
#print "# Running data"
global hypotheses
data_size = args[0]
p_representation = defaultdict(
int) # how often do you get the right representation
p_response = defaultdict(int) # how often do you get the right response?
p_representation_literal = defaultdict(
int) # how often do you get the right representation
p_response_literal = defaultdict(
int) # how often do you get the right response?
p_representation_presup = defaultdict(
int) # how often do you get the right representation
p_response_presup = defaultdict(
int) # how often do you get the right response?
#print "# Generating data"
data = generate_data(data_size)
# recompute these
#print "# Computing posterior"
#[ x.unclear_functions() for x in hypotheses ]
[x.compute_posterior(data) for x in hypotheses]
# normalize the posterior in fs
#print "# Computing normalizer"
Z = logsumexp([x.posterior_score for x in hypotheses])
# and output the top hypotheses
qq = FiniteBestSet(max=True, N=25)
for h in hypotheses:
qq.push(h, h.posterior_score) # get the tops
for i, h in enumerate(qq.get_all(sorted=True)):
for w in h.all_words():
fprintn(8,
data_size,
i,
w,
h.posterior_score,
q(h.value[w]),
f=options.OUT_PATH + "-hypotheses." + str(get_rank()) +
".txt")
# and compute the probability of being correct
#print "# Computing correct probability"
for h in hypotheses:
hstr = str(h)
#print data_size, len(data), exp(h.posterior_score), correct[ str(h)+":"+w ]
for w in words:
p = exp(h.posterior_score - Z)
key = w + ":" + hstr
p_representation[w] += p * (agree_pct[key] == 1.)
p_representation_presup[w] += p * (
agree_pct_presup[key] == 1.
) # if we always agree with the target, then we count as the right rep.
p_representation_literal[w] += p * (agree_pct_literal[key] == 1.)
# and just how often does the hypothesis agree?
p_response[w] += p * agree_pct[key]
p_response_presup[w] += p * agree_pct_presup[key]
p_response_literal[w] += p * agree_pct_literal[key]
#print "# Outputting"
for w in words:
fprintn(10,
str(get_rank()),
q(w),
data_size,
p_representation[w],
p_representation_presup[w],
p_representation_literal[w],
p_response[w],
p_response_presup[w],
p_response_literal[w],
f=options.OUT_PATH + "-stats." + str(get_rank()) + ".txt")
return 0
from LOTlib.Hypotheses.LOTHypothesis import LOTHypothesis
from LOTlib.Inference.Samplers.MetropolisHastings import mh_sample
from LOTlib.Examples.Quantifier.Model import *
ALPHA = 0.9
SAMPLES = 100000
DATA_SIZE = 1000
if __name__ == "__main__":
## sample the target data
data = generate_data(DATA_SIZE)
W = 'every'
# Now to use it as a LOTHypothesis, we need data to have an "output" field which is true/false for whether its the target word. This is then used by LOTHypothesis.compute_likelihood to see if we match or not with whether a word was said (ignoring the other words -- that's why its a pseudolikelihood)
for di in data:
di.output = (di.word == W)
#print (di.word == W)
FBS = FiniteBestSet(max=True, N=100)
H = LOTHypothesis(grammar, args=['A', 'B', 'S'], ALPHA=ALPHA)
# Now just run the sampler with a LOTHypothesis
for s in mh_sample(H, data, SAMPLES, skip=10):
#print s.lp, "\t", s.prior, "\t", s.likelihood, "\n", s, "\n\n"
FBS.push(s, s.lp)
for k in reversed(FBS.get_all(sorted=True)):
print k.lp, k.prior, k.likelihood, k
def make_h0(value=None):
return GaussianLOTHypothesis(grammar, value=value)
if __name__ == "__main__":
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# the running function
def run(*args):
# starting hypothesis -- here this generates at random
h0 = GaussianLOTHypothesis(grammar,
prior_temperature=PRIOR_TEMPERATURE)
# We store the top 100 from each run
pq = FiniteBestSet(100, max=True, key="posterior_score")
pq.add(mh_sample(h0, data, STEPS, skip=SKIP))
return pq
finitesample = FiniteBestSet(max=True) # the finite sample of all
results = map(run, [[None]] * CHAINS) # Run on a single core
finitesample.merge(results)
## and display
for r in finitesample.get_all(decreasing=False, sorted=True):
print r.posterior_score, r.prior, r.likelihood, qq(str(r))
fbs.add(h, h.posterior_score)
return fbs
## ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
### MPI map
## ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
from SimpleMPI.MPI_map import MPI_map, is_master_process
allret = MPI_map(run, map(lambda x: [x], [0.01, 0.1, 1.0] * 100 ))
if is_master_process():
allfbs = FiniteBestSet(max=True)
allfbs.merge(allret)
H = allfbs.get_all()
for h in H:
h.likelihood_temperature = 0.01 # on what set of data we want?
h.compute_posterior(data)
# show the *average* ll for each hypothesis
for h in sorted(H, key=lambda h: h.posterior_score):
print h.posterior_score, h.prior, h.likelihood, h.likelihood_temperature
print h
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
## Play around with some different inference schemes
from Data import data
from Grammar import grammar
from Utilities import make_h0
def run(*args):
"""The running function."""
# starting hypothesis -- here this generates at random
h0 = GaussianLOTHypothesis(grammar)
# We store the top 100 from each run
pq = FiniteBestSet(N=100, max=True, key="posterior_score")
pq.add(MHSampler(h0, data, STEPS, skip=SKIP))
return pq
if __name__ == "__main__":
CHAINS = 10
STEPS = 10000000
SKIP = 0
finitesample = FiniteBestSet(max=True) # the finite sample of all
results = map(run, [ [None] ] * CHAINS ) # Run on a single core
finitesample.merge(results)
## and display
for r in finitesample.get_all(decreasing=False, sorted=True):
print r.posterior_score, r.prior, r.likelihood, qq(str(r))
def novelty_search(h0s, data, grammar, props=10, novelty_advantage=100): """ Search through hypotheses, maintaining a queue of good ones. We propose to ones based on their posterior and how much they've been proposed to in the past. See heapweight(h) below -- it determines how we trade off posterior and prior search there... SO: You are searched further if you are good, and in a "novel" part of the space TODO: We could make this track the performance of proposals from a given hypothesis? """ novelty = defaultdict(float) # last time we proposed here, what proportion were novel? If we haven't done any, set to 1.0 froms = defaultdict(int) # how many times did we propose from this? tos = defaultdict(int) # how many to this? FS = FiniteBestSet(N=10) # When we add something to the heap, what weight does it have? # This should prefer high log probability, but also it should # keep us from looking at things too much def heapweight(h): return -h.lp - novelty[h]*novelty_advantage openset = [] for h0 in h0s: if h0 not in novelty: h0.compute_posterior(data) heapq.heappush( openset, (heapweight(h0),h0) ) novelty[h0] = 1.0 # treat as totally novel FS.add(h0, h0.lp) while not LOTlib.SIG_INTERRUPTED: lph, h = heapq.heappop(openset) froms[h] += 1 #proposals_from[h] += props print "\n" print len(openset), "\t", h.lp, "\t", heapweight(h), "\t", novelty[h], "\t", froms[h], tos[h], "\t", q(h) for x in FS.get_all(sorted=True): print "\t", x.lp, "\t", heapweight(x), "\t", novelty[x], "\t", q(get_knower_pattern(h)), "\t", froms[x], tos[x],"\t", q(x) # Store all together so we know who to update (we make their novelty the same as their parent's) proposals = [ h.propose()[0] for i in xrange(props) ] new_proposals = [] # which are new? novelprop = 0 for p in proposals: if p not in novelty: p.compute_posterior(data) FS.add(p, p.lp) novelty[p] = "ERROR" # just keep track -- should be overwritten later novelprop += 1 new_proposals.append(p) tos[p] += 1 novelty[h] = float(novelprop) / float(props) # use the novelty from the parent for p in new_proposals: novelty[p] = random() * novelty[h] heapq.heappush(openset, (heapweight(p), p) ) # and put myself back on the heap, but with the new proposal numbers heapq.heappush(openset, (heapweight(h), h) )
Run inference on each target concept and save the output
"""
import pickle
from LOTlib import break_ctrlc
from LOTlib.FiniteBestSet import FiniteBestSet
from LOTlib.Inference.Samplers.MetropolisHastings import MHSampler
from Model import *
from TargetConcepts import TargetConcepts
NDATA = 20 # How many data points for each function?
NSTEPS = 100000
BEST_N = 500 # How many from each hypothesis to store
# Where we keep track of all hypotheses (across concepts)
all_hypotheses = FiniteBestSet()
if __name__ == "__main__":
# Now loop over each target concept and get a set of hypotheses
for i, f in enumerate(TargetConcepts):
# Set up the hypothesis
h0 = make_hypothesis()
# Set up some data
data = make_data(NDATA, f)
# Now run some MCMC
fs = FiniteBestSet(N=BEST_N, key="posterior_score")
fs.add(break_ctrlc(MHSampler(h0, data, steps=NSTEPS, trace=False)))
## ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
### MPI map
## ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
try:
from SimpleMPI.MPI_map import MPI_map, is_master_process
except ImportError:
MPI_map = map
is_master_process = lambda: True
allret = MPI_map(run, map(lambda x: [x], [0.01, 0.1, 1.0] * 100 ))
if is_master_process():
allfbs = FiniteBestSet(max=True)
allfbs.merge(allret)
H = allfbs.get_all()
for h in H:
h.likelihood_temperature = 0.01 # on what set of data we want?
h.compute_posterior(data)
# show the *average* ll for each hypothesis
for h in sorted(H, key=lambda h: h.posterior_score):
print h.posterior_score, h.prior, h.likelihood, h.likelihood_temperature
print h
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
## Play around with some different inference schemes
"""
import pickle
from LOTlib import lot_iter
from LOTlib.Hypotheses.LOTHypothesis import LOTHypothesis
from LOTlib.FiniteBestSet import FiniteBestSet
from LOTlib.Inference.MetropolisHastings import MHSampler
from Model import *
NDATA = 50 # How many total data points?
NSTEPS = 10000
BEST_N = 100 # How many from each hypothesis to store
OUTFILE = "hypotheses.pkl"
# Where we keep track of all hypotheses (across concepts)
all_hypotheses = FiniteBestSet()
if __name__ == "__main__":
# Now loop over each target concept and get a set of hypotheses
for i, f in enumerate(TARGET_CONCEPTS):
# Set up the hypothesis
h0 = LOTHypothesis(grammar, start='START', args=['x'])
# Set up some data
data = generate_data(NDATA, f)
# Now run some MCMC
fs = FiniteBestSet(N=BEST_N, key="posterior_score")
fs.add(lot_iter(MHSampler(h0, data, steps=NSTEPS, trace=False)))
data = generate_data(data_size)
# starting hypothesis -- here this generates at random
learner = GriceanQuantifierLexicon(make_my_hypothesis, my_weight_function)
# We will defautly generate from null the grammar if no value is specified
for w in target.all_words(): learner.set_word(w)
# populate the finite sample by running the sampler for this many steps
for x in mh_sample(learner, data, SAMPLES, skip=0):
hypset.push(x, x.posterior_score)
return hypset
if __name__ == "__main__":
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# MPI interface
# Map. SimpleMPI will use a normal MAP if we are not running in MPI
allret = MPI_map(run, map(lambda x: [x], DATA_AMOUNTS * CHAINS)) # this many chains
## combine into a single hypothesis set and save
outhyp = FiniteBestSet(max=True)
for r in allret:
print "# Merging ", len(r)
outhyp.merge(r)
import pickle
pickle.dump(outhyp, open(OUT_PATH, 'w'))
