def search(decode, state, cand):
bos_ind = [word_inds[bos]]
pad_bos = pad_sequences([bos_ind],
maxlen=seq_len,
padding='post',
truncating='post')
pad_bos = add_buf(pad_bos, sym=False)
probs = decode.predict([pad_bos, state])[0][0]
max_probs, max_inds = check(probs, cand, keep_eos=False)
sent2s, log_sums = [bos] * cand, np.log(max_probs)
fin_sent2s, fin_logs = list(), list()
next_words, count = [ind_words[ind] for ind in max_inds], 1
while cand > 0:
log_mat, ind_mat = list(), list()
count = count + 1
for i in range(cand):
sent2s[i] = ' '.join([sent2s[i], next_words[i]])
seq2 = word2ind.texts_to_sequences([sent2s[i]])[0]
pad_seq2 = pad_sequences([seq2],
maxlen=seq_len,
padding='post',
truncating='post')
pad_seq2 = add_buf(pad_seq2, sym=False)
step = min(count - 1, seq_len - 1)
probs = decode.predict([pad_seq2, state])[0][step]
max_probs, max_inds = check(probs, cand, keep_eos=True)
max_logs = np.log(max_probs) + log_sums[i]
log_mat.append(max_logs)
ind_mat.append(max_inds)
max_logs = -np.sort(-np.array(log_mat), axis=None)[:cand]
next_sent2s, next_words, log_sums = list(), list(), list()
for log in max_logs:
args = np.where(log_mat == log)
sent_arg, ind_arg = int(args[0][0]), int(args[1][0])
next_word = ind_words[ind_mat[sent_arg][ind_arg]]
if next_word != eos and count < max_len:
next_words.append(next_word)
next_sent2s.append(sent2s[sent_arg])
log_sums.append(log)
else:
cand = cand - 1
fin_sent2s.append(sent2s[sent_arg])
fin_logs.append(log / count)
sent2s = next_sent2s
max_arg = np.argmax(np.array(fin_logs))
return fin_sent2s[max_arg][2:]
def predict(text, name, mode):
sent1 = clean(text)
seq1 = word2ind.texts_to_sequences([sent1])[0]
pad_seq1 = pad_sequences([seq1],
maxlen=seq_len,
padding='pre',
truncating='pre')
pad_seq1 = add_buf(pad_seq1, sym=True)
encode = map_item(name + '_encode', models)
state = encode.predict(pad_seq1)
decode = map_item(name + '_decode', models)
func = map_item(mode, funcs)
return func(decode, state, cand=3)
def predict(text, name):
seq = word2ind.texts_to_sequences([text])[0]
pad_seq = pad_sequences([seq], maxlen=seq_len)
if name == 'cnn':
pad_seq = add_buf(pad_seq)
model = map_item(name, models)
probs = model.predict(pad_seq)[0]
bound = min(len(text), seq_len)
preds = np.argmax(probs, axis=1)[-bound:]
if __name__ == '__main__':
pairs = list()
for word, pred in zip(text, preds):
pairs.append((word, ind_labels[pred]))
return pairs
else:
return preds
def sample(decode, state, cand):
sent2 = bos
next_word, count = '', 0
while next_word != eos and count < max_len:
count = count + 1
sent2 = ' '.join([sent2, next_word])
seq2 = word2ind.texts_to_sequences([sent2])[0]
pad_seq2 = pad_sequences([seq2],
maxlen=seq_len,
padding='post',
truncating='post')
pad_seq2 = add_buf(pad_seq2, sym=False)
step = min(count - 1, seq_len - 1)
probs = decode.predict([pad_seq2, state])[0][step]
max_probs, max_inds = check(probs, cand, keep_eos=True)
if max_inds[0] == word_inds[eos]:
next_word = eos
else:
max_probs = max_probs / np.sum(max_probs)
next_word = ind_words[choice(max_inds, p=max_probs)]
return sent2[3:]