def training_job(corpus_fname, k, p, seed, static, dev_fname, model_out_prefix):
sents = [s for s in io.conll_to_sents(file(corpus_fname)) if isprojective.is_projective(s)]
print "training ",corpus_fname,k,p,seed,len(sents)
explore = ExplorePolicy(k,p)
TRAIN_OUT_FILE = "%s-ef.kps-k%s-p%s-seed%s" % (model_out_prefix, k, p, seed)
if static:
TRAIN_OUT_FILE = "%s-ef.kps-static-seed%s" % (model_out_prefix, seed)
explore=None
model = Model("features/znp.py", "%s.weights" % TRAIN_OUT_FILE)
model.save("%s.model" % TRAIN_OUT_FILE)
random.seed(seed)
train(sents, model, dev=None, ITERS=20,save_every=None,explore_policy=explore,shuffle_sents=True)
print "training of",corpus_fname,k,p,seed,"done"
print "parsing"
parsed = parse_corpus(dev_fname, TRAIN_OUT_FILE + ".weights.FINAL", "features/znp.py")
outf = file(TRAIN_OUT_FILE + ".dev.parsed","w")
for sent in parsed:
io.out_conll(sent, outf, parent='pparent',prel='pprel')
uas,las,complete = eval(parsed)
puas,plas,complete = eval(parsed,ignore_punct=True)
outf.close()
outf = file(TRAIN_OUT_FILE + ".dev.scores","w")
print >> outf, "UAS:",uas,"LAS:",las,"NP_UAS:",puas,"NP_LAS:",plas
outf.close()
print "deleting"
os.unlink(TRAIN_OUT_FILE + ".weights.FINAL")
os.unlink(TRAIN_OUT_FILE + ".model")
def test_bu_labeler(fname, sents, fext=AnEdgeLabelFeatureExtractor()):
labeler = BottomupSimpleSentenceLabeler(Labeler.load(fname), fext)
stime = time.time()
for sent in sents:
labeler.label(sent, par='parent', prelout='pprel')
io.out_conll(sent, prel='pprel')
print >> sys.stderr, time.time() - stime
def training_job(corpus_fname, k, p, seed, static, dev_fname,
model_out_prefix):
from training import online_greedy_train
sents = [
s for s in io.conll_to_sents(file(corpus_fname))
if isprojective.is_projective(s)
]
print "training ", corpus_fname, k, p, seed, len(sents)
labels = set()
for sent in sents:
for tok in sent:
if tok['prel'] == '_': tok['prel'] = 'dep'
#tok['prel'] = 'dep'
labels.add(tok['prel'])
oracle = ArcHybridStaticOracle() if static else ArcHybridDynamicOracle()
explore = None if static else ExplorePolicy(k, p)
print "start"
feature_extractor = features.extractors.get("hybrid.1")
action_map, params = online_greedy_train(
sents,
transition_system=ArcHybridState,
oracle=oracle,
feature_extractor=feature_extractor,
labels=labels,
iterations=15,
explore_policy=ExplorePolicy(k, p),
random_seed=seed,
shuffle_corpus=True)
print "end"
params.finalize()
TRAIN_OUT_FILE = "%s-hybrid-k%s-p%s-seed%s" % (model_out_prefix, k, p,
seed)
if static:
TRAIN_OUT_FILE = "%s-hybrid-static-seed%s" % (model_out_prefix, seed)
params.dump(file(TRAIN_OUT_FILE, "w"), sparse=True)
pickle.dump(action_map, file(TRAIN_OUT_FILE + ".amap", "w"))
print "training of", corpus_fname, k, p, seed, "done"
print "parsing"
parsed = parse_corpus(dev_fname, TRAIN_OUT_FILE, feature_extractor,
ArcHybridState)
print "writing"
outf = file(TRAIN_OUT_FILE + ".dev.parsed", "w")
for sent in parsed:
io.out_conll(sent, outf, parent='pparent', prel='pprel')
uas, las, complete = eval(parsed)
puas, plas, complete = eval(parsed, ignore_punct=True)
outf.close()
outf = file(TRAIN_OUT_FILE + ".dev.scores", "w")
print >> outf, "UAS:", uas, "LAS:", las, "NP_UAS:", puas, "NP_LAS:", plas
print "UAS:", uas, "LAS:", las, "NP_UAS:", puas, "NP_LAS:", plas
outf.close()
print "deleting"
os.unlink(TRAIN_OUT_FILE)
os.unlink(TRAIN_OUT_FILE + ".amap")
def test_labeler(fname,
sents,
fext=AnEdgeLabelFeatureExtractor(),
guides=_dummyguides()):
labeler = SimpleSentenceLabeler(Labeler.load(fname), fext)
stime = time.time()
for sent, sguide in zip(sents, guides):
labeler.label(sent, par='parent', prelout='pprel', sent_guides=sguide)
io.out_conll(sent, prel='pprel')
print >> sys.stderr, "time:", time.time() - stime
def conll_out(sent, extra=None, outf=sys.stdout):
from pio import io
for rel, par, chl in extra:
newchl = copy.copy(chl)
newchl['tag'] = 'COIDX_%s' % chl['id']
newchl['id'] = len(sent) + 1
newchl['pparent'] = par['id']
newchl['prel'] = rel
sent.append(newchl)
io.out_conll(sent, parent='pparent', out=outf)
def test():
'''
Test Model
'''
logging.info("test ...")
featExt = extractors.get(FLAGS.feature_extarctor)
p = ArcEagerParser(
MLActionDecider(ml.MulticlassModel(FLAGS.model), featExt))
good = 0.0
bad = 0.0
complete = 0.0
# main test loop
reals = set()
preds = set()
with open(FLAGS.test_results, "w") as fout:
sents = io.transform_conll_sents(FLAGS.test_data,
FLAGS.only_projective, FLAGS.unlex)
for i, sent in enumerate(sents):
sgood = 0.0
sbad = 0.0
mistake = False
sys.stderr.write("%s %s %s\n" % ("@@@", i, good /
(good + bad + 1)))
try:
d = p.parse(sent)
except MLTrainerWrongActionException:
# this happens only in "early update" parsers, and then we just go on to
# the next sentence..
continue
sent = d.annotate_allow_none(sent)
for tok in sent:
if FLAGS.ignore_punc and tok['form'][0] in "`',.-;:!?{}":
continue
reals.add((i, tok['parent'], tok['id']))
preds.add((i, tok['pparent'], tok['id']))
if tok['pparent'] == -1:
continue
if tok['parent'] == tok['pparent'] or tok['pparent'] == -1:
good += 1
sgood += 1
else:
bad += 1
sbad += 1
mistake = True
if FLAGS.unlex:
io.out_conll(sent, parent='pparent', form='oform')
else:
io.out_conll(sent, parent='pparent', form='form')
if not mistake:
complete += 1
# sys.exit()
logging.info("test result: sgood[%s], sbad[%s]", sgood, sbad)
if sgood > 0.0 and sbad > 0.0:
fout.write("%s\n" % (sgood / (sgood + sbad)))
logging.info("accuracy: %s", good / (good + bad))
logging.info("complete: %s", complete / len(sents))
preds = set([(i, p, c) for i, p, c in preds if p != -1])
logging.info("recall: %s",
len(preds.intersection(reals)) / float(len(reals)))
logging.info("precision: %s",
len(preds.intersection(reals)) / float(len(preds)))
logging.info("assigned: %s", len(preds) / float(len(reals)))
now = time.time()
for i, sent in enumerate(sents):
sgood = 0.0
sbad = 0.0
mistake = False
print >> sys.stdout, "%s %s %s\n" % ("@@@", i, good / (good + bad + 1))
#io.out_conll(sent)
try:
d = p.parse(sent)
except MLTrainerWrongActionException:
# this happens only in "early update" parsers, and then we just go on to
# the next sentence..
print "WTF"
continue
sent = d.annotate(sent)
io.out_conll(sent, parent='parent')
io.out_conll(sent, parent='pparent')
for tok in sent:
#print tok['parent'],tok['pparent']
if FLAGS.ignore_punc and tok['form'][0] in "`',.-;:!?{}": continue
reals.add((i, tok['parent'], tok['id']))
preds.add((i, tok['pparent'], tok['id']))
if FLAGS.ignore_punc and tok['prel'] == 'punct': continue
if tok['pparent'] == -1: continue
if tok['parent'] == tok['pparent'] or tok['pparent'] == -1:
if tok['prel'] == tok['pprel']: lgood += 1
else: print "badl:", tok['prel'], tok['pprel']
good += 1
sgood += 1
else:
bad += 1