def get_corpora(lang, num_train=500000, num_test=10000, distributed=False):
full_corpus = corpora.get_corpus(lang, word_boundaries=True)
# A list of (phoneme, precedes_boundary) tuples.
phones_and_boundaries = extract_boundaries(full_corpus)
# Divide into train and test.
train, test = corpora.train_test_split(phones_and_boundaries,
num_train, num_test, mode='end')
# Separate phones from boundary markers.
train_phones, _ = map(list, zip(*train))
test_phones, test_bounds = map(list, zip(*test))
joblib.dump(test_bounds, lang + '_bounds.pkl')
return
# Construct targets and encode phonemes.
train_in, train_out = prepare(train_phones, distributed)
test_in, test_out = prepare(test_phones, distributed)
# Remove the trailing bound to match test_out.
del test_bounds[-1]
assert len(train_in) == len(train_out)
assert len(test_in) == len(test_out) == len(test_bounds)
return (train_in, train_out), (test_in, test_out), test_bounds
def get_corpora(lang, kind, train_len, roc_len=100, bleu_len=100):
if lang == 'toy2':
import pcfg
corpus = (s.split(' ') for s in pcfg.toy2())
else:
corpus = corpora.get_corpus(lang, kind)
train_corpus = [next(corpus) for _ in range(train_len)]
testable = (utt for utt in corpus if 2 < len(utt) < 20)
roc_test_corpus = [next(testable) for _ in range(roc_len)]
producable = (utt for utt in corpus if 2 < len(utt) < 20)
bleu_test_corpus = [next(producable) for _ in range(bleu_len)]
return {'train': train_corpus,
'roc_test': roc_test_corpus,
'bleu_test': bleu_test_corpus}
def get_model(train_len=1000, lang='english', kind='word', **params):
# for testing
model = Numila(**params)
corpus = corpora.get_corpus(lang, kind)
train_corpus = [next(corpus) for _ in range(train_len)]
return model.fit(train_corpus)
