def __call__(self, data):
vec = Vector()
for i, (key, val) in enumerate(data):
splits = val.split("****")
if len(splits) <> 2:
print >>sys.stderr,"skipping sent"
continue
sent, oracle = splits
s2 = sent.replace("\t\t\t", "\n")
o2 = oracle.replace("\t\t\t", "\n")
sent_forest = Forest.load(StringIO(s2), True, lm=None).next()
oracle_forest = Forest.load(StringIO(o2), True, lm=None).next()
assert sent_forest, oracle_forest
#print >>sys.stderr, len(sent_forest)
#print >>sys.stderr, len(oracle_forest)
example_marg, example_partition = fast_inside_outside.collect_marginals(sent_forest, self.weights)
oracle_marg, oracle_partition = fast_inside_outside.collect_marginals(oracle_forest, self.weights)
vec += example_marg - oracle_marg
vec["log_likelihood"] += example_partition-oracle_partition
#vec["log_likelihood"] += example_partition-oracle_partition
self.processed += 1
for feat in vec:
yield feat, vec[feat]
def compute_marginals(self, forest, oracle_forest):
"computes the marginals of a -lm forest"
# print >> logs, "Example TIME %s"%((end - start))
# oracle_bleu, oracle_trans, oracle_fv, _ = oracle_forest.compute_oracle(Vector(), model_weight=0.0, bleu_weight=1.0)
def non_local_scorer(cedge, ders):
hyp = cedge.assemble(ders)
return ((0.0, Vector()), hyp, hyp)
# decoder = CubePruning(MarginalDecoder.FeatureAdder(self.weights), non_local_scorer, 20, 5, find_min=False)
# best = decoder.run(forest.root)
# example_marginals = Vector()
# total = -INF
# for i in range(min(200, len(best))):
# M = max(best[i].score[0], total)
# m = min(best[i].score[0], total)
# total = M + log(1.0 + exp(m - M))
# #print "before"
# print total
# for i in range(min(200, len(best))):
# #print exp(best[i].score[0] -total)
# example_marginals += exp(best[i].score[0] -total) * best[i].score[1]
# #print "after"
# partition = total
# start = time.time()
example_marg, partition = fast_inside_outside.collect_marginals(forest, self.weights)
# end = time.time()
# print >> logs, "marg TIME %s"%((end - start))
# print "Best Log Likelihood %s "%(best[0].score[0] - partition)
# start = time.time()
# oracle_forest, oracle_item = oracle.oracle_extracter(forest, self.weights, 5, 2, extract=1)
# end = time.time()
# print >> logs, "oracle forest %s"%((end - start))
# start = time.time()
oracle_marg, oracle_partition = fast_inside_outside.collect_marginals(oracle_forest, self.weights)
(oracle_best, oracle_subtree, oracle_best_fv) = oracle_forest.bestparse(self.weights, use_min=False)
# end = time.time()
# print >> logs, "oracle TIME %s"%((end - start))
# logs.flush()
# self.write_model("", oracle_marg)
# print "Best Score: %s"% best
# print "Oracle Score: %s"% (self.weights.dot(oracle_fv))
# for i in range(5):
# print "Oracle Trans: %s %s" %(oracle_item[i].full_derivation, oracle_item[i].score)
# print "Oracle BLEU Score: %s"% (forest.bleu.rescore(oracle_item[i].full_derivation))
# forest.bleu.rescore(oracle_subtree)
# print "Oracle Trans: %s %s" %(oracle_subtree, forest.bleu.score_ratio_str())
# print "Best Trans: %s"%best[0].full_derivation
# forest.bleu.rescore(best[0].full_derivation)
# print "Best BLEU Score: %s"% (forest.bleu.score_ratio_str())
# print oracle_partition -partition, oracle_partition, partition
average = 0.0
# for i in range(min(10, len(best))):
# print " Best Trans: %s"%best[i].full_derivation
# forest.bleu.rescore(best[i].full_derivation)
# average += len(best[i].full_derivation.split())
# print " Best BLEU Score: %s"% (forest.bleu.score_ratio_str())
# print " Best Score: %s"% (best[i].score[0])
# print "Average Length %s"%(average / float(i))
# print "Local Difference: %s"%(oracle_partition-partition)
return example_marg, oracle_marg, oracle_partition - partition # log div
