bleu, hyp, fv, edgelist = forest.compute_oracle(weights, 0, 1, store_oracle=True)
##print >> logs, forest.root.oracle_edge
bleu = forest.bleu.rescore(hyp)
mscore = weights.dot(fv)
print >> logs, "moracle\tscore=%.4lf\tbleu+1=%.4lf\tlenratio=%.2lf" % \
(mscore, forest.bleu.fscore(), forest.bleu.ratio()) #, hyp) # don't output oracle
myoraclebleus += forest.bleu.copy()
myscores += mscore
if FLAGS.out:
if FLAGS.suffix is not None:
forest.dump(open("%d.%s" % (i+FLAGS.startid, FLAGS.suffix), "wt"))
else:
forest.dump()
if i % 10 == 0:
if old_edges <> 0:
print >> logs, "%d forests pruned, avg new size: %.1lf %.1lf (survival ratio: %4.1lf%% %4.1lf%%)" % \
(i, total_nodes/i, total_edges/i, \
total_nodes*100./old_nodes, total_edges*100./old_edges)
print >> logs, "overall 1-best oracle bleu = %s score = %.4lf" \
% (onebestbleus.score_ratio_str(), onebestscores/i)
if FLAGS.oracle:
print >> logs, "overall my oracle bleu = %s score = %.4lf" \
% (myoraclebleus.score_ratio_str(), myscores/i)
def main():
weights = Model.cmdline_model()
lm = Ngram.cmdline_ngram()
LMState.init(lm, weights)
decoder = Decoder()
tot_bleu = Bleu()
tot_score = 0.
tot_time = 0.
tot_len = tot_fnodes = tot_fedges = 0
tot_steps = tot_states = tot_edges = tot_stacks = 0
for i, forest in enumerate(Forest.load("-", is_tforest=True, lm=lm), 1):
t = time.time()
best, final_items = decoder.beam_search(forest, b=FLAGS.beam)
score, trans, fv = best.score, best.trans(), best.get_fvector()
t = time.time() - t
tot_time += t
tot_score += score
forest.bleu.rescore(trans)
tot_bleu += forest.bleu
fnodes, fedges = forest.size()
tot_len += len(forest.sent)
tot_fnodes += fnodes
tot_fedges += fedges
tot_steps += decoder.max_step
tot_states += decoder.num_states
tot_edges += decoder.num_edges
tot_stacks += decoder.num_stacks
print >> logs, ("sent %d, b %d\tscore %.4f\tbleu+1 %s" + \
"\ttime %.3f\tsentlen %-3d fnodes %-4d fedges %-5d\tstep %d states %d edges %d stacks %d") % \
(i, FLAGS.beam, score,
forest.bleu.score_ratio_str(), t, len(forest.sent), fnodes, fedges,
decoder.max_step, decoder.num_states, decoder.num_edges, decoder.num_stacks)
if FLAGS.k > 1 or FLAGS.forest:
lmforest = best.toforest(forest)
if FLAGS.forest:
lmforest.dump()
if FLAGS.k > 1:
lmforest.lazykbest(FLAGS.k, weights=weights)
klist = lmforest.root.klist
if not FLAGS.mert:
for j, (sc, tr, fv) in enumerate(klist, 1):
print >> logs, "k=%d score=%.4f fv=%s\n%s" % (j, sc, fv, tr)
else:
klist = [(best.score, best.trans(), best.get_fvector())]
if FLAGS.mert: # <score>... <hyp> ...
print >> logs, '<sent No="%d">' % i
print >> logs, "<Chinese>%s</Chinese>" % " ".join(forest.cased_sent)
for sc, tr, fv in klist:
print >> logs, "<score>%.3lf</score>" % sc
print >> logs, "<hyp>%s</hyp>" % tr
print >> logs, "<cost>%s</cost>" % fv
print >> logs, "</sent>"
if not FLAGS.forest:
print trans
print >> logs, "avg %d sentences, first pass score: %.4f, bleu: %s" % \
(i, decoder.firstpassscore/i, decoder.firstpassbleus.score_ratio_str())
print >> logs, ("avg %d sentences, b %d\tscore %.4lf\tbleu %s\ttime %.3f" + \
"\tsentlen %.1f fnodes %.1f fedges %.1f\tstep %.1f states %.1f edges %.1f stacks %.1f") % \
(i, FLAGS.beam, tot_score/i, tot_bleu.score_ratio_str(), tot_time/i,
tot_len/i, tot_fnodes/i, tot_fedges/i,
tot_steps/i, tot_states/i, tot_edges/i, tot_stacks/i)
print >> logs, LMState.cachehits, LMState.cachemiss
