def decode(self, sess, tokens, chars, length, topK=5):
"""
score: [seq_len, num_tags]
transition_params: [num_tags, num_tags]
"""
score, trans_params = sess.run(
[self.unary_potentials, self.trans_params], {
self.tokens: tokens,
self.chars: chars,
self.dropout: 1.0
})
score = np.squeeze(score, 0)
score = score[:length, :]
'''
viterbi, viterbi_score = tf.contrib.crf.viterbi_decode(
score,
trans_params
)
print("{:<20} {}".format(viterbi_score, viterbi))
'''
viterbi, viterbi_score = viterbi_decode_topk(score, trans_params, topK)
for a, b in zip(viterbi_score, viterbi):
print("{:<20} {}".format(a, b))
return viterbi, viterbi_score
def decode(self, sess, tokens, chars, topK=5):
"""
score: [seq_len, num_tags]
transition_params: [num_tags, num_tags]
"""
lower_tokens = np.char.lower(tokens)
input_feed = {
self.tokens: lower_tokens,
self.chars: chars,
self.dropout: 1.0
}
if self.elmo_bilm and self.elmo_batcher:
input_feed[self.elmo_p] = self.elmo_batcher.batch_sentences(
[sx.tolist() for sx in tokens])
scores, trans_params, lengths = sess.run(
[self.unary_potentials, self.trans_params, self.length],
input_feed)
out = []
for score, length in zip(scores, lengths):
score = score[:length, :]
viterbi, viterbi_score = viterbi_decode_topk(
score, trans_params, topK)
out.append(viterbi)
'''
viterbi, viterbi_score = tf.contrib.crf.viterbi_decode(
score,
trans_params
)
print("{:<20} {}".format(viterbi_score, viterbi))
'''
'''
viterbi, viterbi_score = viterbi_decode_topk(
score,
trans_params,
topK
)
'''
# for a, b in zip(viterbi_score, viterbi):
# print("{:<20} {}".format(a, b))
# return viterbi, viterbi_score
return out