def decode(self, forest, early_stop=False):
# decoding
oracle_bleu, oracle_trans, oracle_fv, _ = forest.compute_oracle(Vector(), model_weight=0.0, bleu_weight=1.0)
(score, best_trans, best_fv) = self.decoder.beam_search(forest, b=FLAGS.beam)
# self.write_model("", best_fv)
best_lm = self._add_back_language_model(best_trans)
# best_lm1 = self._add_back_uni_language_model(best_trans)
# assert best_lm == best_fv["lm"], "%s, %s "% (best_lm, best_fv["lm"])
# assert best_lm1 == best_fv["lm1"], "%s, %s "% (best_lm1, best_fv["lm1"])
oracle_fv["lm"] = self._add_back_language_model(oracle_trans)
# print "oracle", oracle_fv["lm1"]
# oracle_fv["lm1"] = self._add_back_uni_language_model(oracle_trans)
# print "test score " + str(score)
# print "dot test " + str(self.weights.dot(best_fv))
test_bleu = forest.bleu.rescore(best_trans)
oracle_bleu = forest.bleu.rescore(oracle_trans)
# print "oracle test " + str(self.weights.dot(oracle_fv))
# print best_trans
# print oracle_trans
# best_bleu = test_bleu
# oracle_fv = best_fv
# for k, (sc, trans, fv) in enumerate(forest.root.hypvec, 1):
# hyp_bleu = forest.bleu.rescore(trans)
# if hyp_bleu > best_bleu:
# oracle_fv = fv
# best_bleu = hyp_bleu
print >> sys.stderr, "Test: %s \n Oracle: %s" % (best_trans, oracle_trans)
print >> sys.stderr, "BLEU Test: %s | Oracle: %s" % (test_bleu, oracle_bleu)
print >> sys.stderr, "MODEL Test: %s | Oracle: %s" % (self.weights.dot(best_fv), self.weights.dot(oracle_fv))
# delta_feats = self.clip(self.trim(-oracle_fv + best_fv))
delta_feats = self.trim(-oracle_fv + best_fv)
# self.write_model("",best_fv)
# print " ----------------------"
# self.write_model("",oracle_fv)
#
# self.write_model("", delta_feats)
# print " ----------------------"
# self.write_model("", self.weights)
# self.write_model("", best_fv)
# self.write_model("", oracle_fv)
test_bleu = forest.bleu.rescore((best_trans))
return (forest.bleu, delta_feats)
def decode(self, forest, verbose=False):
# decoding
# 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)
# for i in range(min(10, len(best))):
# print " Best Trans: %s"%best[i].full_derivation
# forest.bleu.rescore(best[i].full_derivation)
# print " Best BLEU Score: %s"% (forest.bleu.score_ratio_str())
# print " Best Score: %s"% (best[i].score[0])
# test_bleu = forest.bleu.rescore((best[0].full_derivation))
# oracle_forest, oracle_item = oracle.oracle_extracter(forest, self.weights, None, 5, 2, extract=1)
(score, subtree, fv) = forest.bestparse(self.weights, use_min=False)
print "textlenght", fv["Basic/text-length"], self.weights["Basic/text-length"], score
if verbose:
print "Ref Tran %s" % forest.refs
(oracle_score, oracle_subtree, oracle_fv, _) = forest.compute_oracle(self.weights)
print "Oracle: " + oracle_subtree
# oracle_forest, oracle_item = oracle.oracle_extracter(forest, self.weights, None, 100, 2, extract=100)
# (oracle_forest_score , oracle_forest_subtree, oracle_forest_fv) = oracle_forest.bestparse(self.weights, use_min=False)
# example_marg, partition = fast_inside_outside.collect_marginals(forest, self.weights)
# oracle_marg, oracle_partition = fast_inside_outside.collect_marginals(oracle_forest, self.weights)
# print "Oracle forest likelihood: ",oracle_partition - partition
# print "Oracle Forest Score: ", oracle_forest_score
# print "Oracle Forest Results: ", oracle_forest_subtree
# print "Oracle Forest Bleu: ", forest.bleu.rescore(oracle_forest_subtree)
# parse_diff(self.weights,fv, oracle_forest_fv)
print "Best: ", subtree
print "Best Score: ", (score, self.weights.dot(oracle_fv))
print "Local Score: ", forest.bleu.rescore(subtree)
# (score, best_trans, best_fv) = self.decoder.beam_search(forest, b=FLAGS.beam)
print "\n\n\n"
# print delta_feats
forest.bleu.rescore(subtree)
return (forest.bleu, None)
