def string_kld(target,
hypothesis,
samples=1000,
verbose=False,
max_length=math.inf,
seed=None):
### sample n trees from target.
if seed == None:
rng = numpy.random.RandomState()
else:
rng = numpy.random.RandomState(seed)
inside_target = wcfg.InsideComputation(target)
inside_hypothesis = wcfg.InsideComputation(hypothesis)
sampler = wcfg.Sampler(target, random=rng)
total = 0.0
n = 0
for i in range(samples):
t = sampler.sample_tree()
s = utility.collect_yield(t)
if len(s) <= max_length:
n += 1
lp = inside_target.inside_log_probability(s)
lq = inside_hypothesis.inside_log_probability(s)
if verbose:
print("Sample %d %s, target %f, hypothesis %f" %
(i, t, lp, lq))
total += lp - lq
return total / n
def __init__(self, target_pcfg):
self.target_pcfg = target_pcfg
self.terminals = set(target_pcfg.terminals)
self.te = target_pcfg.terminal_expectations()
self.pe = target_pcfg.production_expectations()
self.sampler = wcfg.Sampler(self.target_pcfg)
self.renyi = RENYI
def __init__(self, target_pcfg):
self.target_pcfg = target_pcfg
self.terminals = set(target_pcfg.terminals)
self.te = target_pcfg.terminal_expectations()
self.pe = target_pcfg.production_expectations()
self.sampler = wcfg.Sampler(self.target_pcfg)
self.insider = wcfg.InsideComputation(self.target_pcfg)
self.contexts_map = {}
def monte_carlo(target, f, samples, max_length, seed=None):
## f is called on each element
if seed == None:
rng = numpy.random.RandomState()
else:
rng = numpy.random.RandomState(seed)
sampler = wcfg.Sampler(target, random=rng)
for i in range(samples):
t = sampler.sample_tree()
s = utility.collect_yield(t)
if len(s) <= max_length:
f(t, i)
def estimate_bijection(target_pcfg,
hypothesis_pcfg,
samples=1000,
seed=None,
max_length=math.inf,
verbose=False):
## Essential assumption
assert len(target_pcfg.nonterminals) == len(hypothesis_pcfg.nonterminals)
if seed == None:
rng = numpy.random.RandomState()
else:
rng = numpy.random.RandomState(seed)
sampler = wcfg.Sampler(target_pcfg, random=rng)
insider = wcfg.InsideComputation(hypothesis_pcfg)
n = len(target_pcfg.nonterminals)
c = defaultdict(Counter)
for _ in range(samples):
tree = sampler.sample_tree()
s = utility.collect_yield(tree)
if len(s) <= max_length:
try:
learned = insider.bracketed_viterbi_parse(tree)
collect_nonterminal_pairs(tree, learned, c)
except utility.ParseFailureException:
print("Failed", s)
#print(c)
maximum_value = max(max(c2.values()) for nt, c2 in c.items())
#print(maximum_value)
## Now use the Hungarian algorithm
cost_matrix = np.zeros((n, n))
target_list = list(target_pcfg.nonterminals)
hypothesis_list = list(hypothesis_pcfg.nonterminals)
for i, a in enumerate(target_list):
for j, b in enumerate(hypothesis_list):
count = c[a][b]
cost_matrix[i, j] = maximum_value - count
# if count == 0:
# cost_matrix[i,j] = maximum_value
# else:
# cost_matrix[i,j] = 1/count ## Maybe normalise so as to maximize something else?
row_ind, col_ind = linear_sum_assignment(cost_matrix)
answer = {}
for i, j in zip(row_ind, col_ind):
answer[target_list[i]] = hypothesis_list[j]
#print(answer)
return answer
def labeled_exact_match(target,
hypothesis,
samples=1000,
max_length=30,
viterbi=False,
verbose=False,
seed=None):
"""
Proportion of trees whose viterbi parse is the same up to a relabeling of the hypothesis tree.
Target has to be a pcfg; hypothesis can be any WCFG.
Identical nonterminals
"""
if seed == None:
rng = numpy.random.RandomState()
else:
rng = numpy.random.RandomState(seed)
sampler = wcfg.Sampler(target, random=rng)
if viterbi:
inside_target = wcfg.InsideComputation(target)
inside_hypothesis = wcfg.InsideComputation(hypothesis)
total = 0.0
n = 0
for i in range(samples):
t = sampler.sample_tree()
s = utility.collect_yield(t)
if len(s) >= max_length:
continue
n += 1
if viterbi:
t = inside_target.viterbi_parse(s)
try:
th = inside_hypothesis.viterbi_parse(s)
#relabeled_tree = utility.relabel_tree(th, ntmap)
relabeled_tree = th
if relabeled_tree == t:
total += 1
elif verbose:
logging.info("Mismatch in trees with parse of %s", s)
print(relabeled_tree)
print(t)
except utility.ParseFailureException as e:
# Treat this as a failure .
print("Parse failure of %s " % s)
return total / n
def sample_contexts(self, nt):
a = self.kernel_map[nt]
iwcfg = self.target_pcfg.intersect(a)
iwcfg.renormalise()
iwcfg.locally_normalise()
asampler = wcfg.Sampler(iwcfg)
contexts = []
for _ in range(maxsamples):
w = asampler.sample_string()
positions = [i for i, b in enumerate(w) if b == a]
assert len(positions) > 0
position = numpy.random.choice(positions)
left = w[:position]
right = w[position + 1:]
contexts.append((tuple(left), tuple(right)))
return contexts
def test(self):
"""
Test the target PCFG.
return a Wcfg, or None if the grammar is not anchored.
"""
ntmap = self.test_if_anchored()
if len(ntmap) == 0:
return None
print(ntmap)
self.nonterminal_map = ntmap
self.kernel_map = {ntmap[a]: a for a in ntmap}
self.kernel = ntmap.values()
print(self.kernel)
self.kernel_grammar = self.make_kernel_grammar()
self.sampler = wcfg.Sampler(self.kernel_grammar)
self.test_all(N_SAMPLES, MAX_SAMPLES)
return self.make_grammar()
def conditional_kld(target,
hypothesis,
samples=1000,
verbose=False,
max_length=math.inf,
seed=None):
"""
Estimate the kld between the conditional probability distributions.
for a given string $w$ D( P(tree|w) | Q(tree|w)).
difference between string KLD and tree KLD.
Target must be a pcfg, hyppthesis can be arbitrary wcfg, not even convergent, with same nonterminals.
"""
if seed == None:
rng = numpy.random.RandomState()
else:
rng = numpy.random.RandomState(seed)
inside_target = wcfg.InsideComputation(target)
inside_hypothesis = wcfg.InsideComputation(hypothesis)
sampler = wcfg.Sampler(target, random=rng)
total = 0.0
n = 0
for i in range(samples):
t = sampler.sample_tree()
s = utility.collect_yield(t)
if len(s) > max_length:
if verbose:
print("Skipping", len(s))
continue
n += 1
ptree = target.log_probability_derivation(t)
pstring = inside_target.inside_log_probability(s)
qtree = hypothesis.log_probability_derivation(t)
qstring = inside_hypothesis.inside_log_probability(s)
total += (ptree - qtree) - (pstring - qstring)
if verbose:
print("%s p(t) = %f, p(w) = %f, q(t) = %f, q(w) = %f" %
(s, ptree, pstring, qtree, qstring))
return total / n
def sample_contexts_smart(self, a, nsamples, maxsamples):
#print("Sampling contexts for ", a)
iwcfg = self.target_pcfg.intersect(a)
iwcfg.renormalise()
iwcfg.locally_normalise()
asampler = wcfg.Sampler(iwcfg)
contexts = set()
for _ in range(maxsamples):
w = asampler.sample_string()
for i, b in enumerate(w):
if a == b:
left = w[:i]
right = w[i + 1:]
contexts.add((tuple(left), tuple(right)))
if len(contexts) >= nsamples:
return contexts
return contexts
help="just output the yield",
action="store_true")
## Other options: control output format, what probs are calculated.
args = parser.parse_args()
mypcfg = wcfg.load_wcfg_from_file(args.inputfilename)
if args.seed:
print("Setting seed to ", args.seed)
prng = RandomState(args.seed)
else:
prng = RandomState()
mysampler = wcfg.Sampler(mypcfg, random=prng)
insider = wcfg.InsideComputation(mypcfg)
with open(args.outputfilename, 'w') as outf:
i = 0
while i < args.n:
tree = mysampler.sample_tree()
# default is string.
s = utility.collect_yield(tree)
if not args.maxlength or len(s) <= args.maxlength:
if not args.omitprobs:
lps = insider.inside_log_probability(s)
lpt = mypcfg.log_probability_derivation(tree)
lpb = insider._bracketed_log_probability(tree)[mypcfg.start]
outf.write("%0.12f %0.12f %0.12f " % (lpt, lpb, lps))
if args.yieldonly: