def predict(self, keywords):
sentences = []
for keyword in keywords:
content_inputs, content_inputs_length, keyword_inputs, keyword_length = prepare_batch_predict_data(
keyword,
previous=sentences,
prev=FLAGS.prev_data,
rev=FLAGS.rev_data,
align=FLAGS.align_data)
predicted_batch = self.model.predict(self.sess, content_inputs,
content_inputs_length,
keyword_inputs,
keyword_length)
predicted_line = predicted_batch[
0] # predicted is a batch of one line
predicted_line_clean = predicted_line[:-1] # remove the end token
predicted_ints = map(
lambda x: x[0], predicted_line_clean
) # Flatten from [time_step, 1] to [time_step]
predicted_sentence = ints_to_sentence(predicted_ints)
if FLAGS.rev_data:
predicted_sentence = predicted_sentence[::-1]
sentences.append(predicted_sentence)
return sentences
def predict(self, keywords):
sentences = []
keywords = self.Normalize(keywords) # add by zjg
for keyword in keywords:
source, source_len = prepare_batch_predict_data(
keyword,
previous=sentences,
prev=FLAGS.prev_data,
rev=FLAGS.rev_data,
align=FLAGS.align_data)
with self.sess.as_default():
with self.graphpre.as_default():
predicted_batch = self.model.predict(
self.sess,
encoder_inputs=source,
encoder_inputs_length=source_len)
predicted_line = predicted_batch[
0] # predicted is a batch of one line
predicted_line_clean = predicted_line[:-1] # remove the end token
predicted_ints = map(
lambda x: x[0], predicted_line_clean
) # Flatten from [time_step, 1] to [time_step]
predicted_sentence = ints_to_sentence(predicted_ints)
if FLAGS.rev_data:
predicted_sentence = predicted_sentence[::-1]
sentences.append(predicted_sentence)
return sentences
def predict_with_couplet(self, keywords):
sentences = []
for keyword in keywords[::2]:
source, source_len = prepare_batch_predict_data(
keyword,
previous=sentences,
prev=FLAGS.prev_data,
rev=FLAGS.rev_data,
align=FLAGS.align_data)
predicted_batch = self.model.predict(
self.sess,
encoder_inputs=source,
encoder_inputs_length=source_len)
# predicted is a batch of one line
predicted_line = predicted_batch[0]
predicted_line_clean = predicted_line[:-1] # remove the end token
# Flatten from [time_step, 1] to [time_step]
predicted_ints = map(lambda x: x[0], predicted_line_clean)
predicted_sentence = ints_to_sentence(predicted_ints)
if FLAGS.rev_data:
predicted_sentence = predicted_sentence[::-1]
sentences.append(predicted_sentence)
# connects to a socket
self.soc = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
self.soc.connect(("127.0.0.1", 12345))
# we must encode the string to bytes
self.soc.send((sentences[0] + predicted_sentence).encode("utf8"))
# the number means how the response can be in bytes
result_bytes = self.soc.recv(4096)
self.soc.close()
# the return will be in bytes, so decode
pair_sentence = result_bytes.decode("utf8")
# pair_sentence = subprocess.check_output(["python3 couplet_gen/seq_learning.py " + predicted_sentence])
sentences.append(pair_sentence)
return sentences
def predict(self, keywords):
sentences = []
for idx, keyword in enumerate(keywords):
source, source_len = prepare_batch_predict_data(
keyword,
previous=sentences,
prev=FLAGS.prev_data,
rev=FLAGS.rev_data,
align=FLAGS.align_data)
# flag_pre = True
# while flag_pre:
predicted_batch = self.model.predict(
self.sess,
encoder_inputs=source,
encoder_inputs_length=source_len)
# print("p",predicted_batch)
predicted_line = predicted_batch[
0] # predicted is a batch of one line
predicted_line_clean = predicted_line[:-1] # remove the end token
predicted_ints = [x[0] for x in predicted_line_clean
] # Flatten from [time_step, 1] to [time_step]
predicted_sentence = ints_to_sentence(predicted_ints)
if FLAGS.rev_data:
predicted_sentence = predicted_sentence[::-1]
# if idx == 0 or idx ==2:
# flag_pre = False
# if idx == 1:
# print("p",predicted_sentence[-1])
# lis1 = RhymeUtil.get_possible_rhyme_categories(ch = predicted_sentence[-1])
# flag_pre = False
# if idx == 3:
# lis3 = RhymeUtil.get_possible_rhyme_categories(ch = predicted_sentence[-1])
# for ch in lis3:
# if ch in lis1:
# flag_pre = False
sentences.append(predicted_sentence)
return sentences
def predict(self, keywords):
sentences = []
sentences_no_space = []
for keyword in keywords:
source, source_len = prepare_batch_predict_data(
keyword,
previous=sentences,
prev=FLAGS.prev_data,
rev=FLAGS.rev_data,
align=FLAGS.align_data,
keep_prev_num=FLAGS.keep_prev_num,
sen_pad_len=FLAGS.sen_pad_len,
keyword_pad_len=FLAGS.keyword_pad_len)
# print 'source: ', source
# print 'source_len: ', source_len
predicted_batch = self.model.predict(
self.sess,
encoder_inputs=source,
encoder_inputs_length=source_len)
# added by lrx, store model in binary format, so Java can call it
# tensor_names = [t.name for op in tf.get_default_graph().get_operations() for t in op.values()]
# for i in tensor_names: print i
# output_graph_def = tf.graph_util.convert_variables_to_constants(self.sess, self.sess.graph_def, output_node_names=['seq2seq/decoder/out_put0000'])
# tf.train.write_graph(output_graph_def, FLAGS.model_dir, 'minimal_graph.proto', as_text=False)
# parse decoding results
if FLAGS.beam_width > 1:
predicted_line = predicted_batch[0]
# print predicted_line
sent_set = []
for i in range(FLAGS.beam_width):
predicted_ints = []
for j in range(len(predicted_line)):
if predicted_line[j][i] == get_vocab_size() - 2:
break
else:
predicted_ints.append(predicted_line[j][i])
# reverse first, then map ints to words
if FLAGS.rev_data:
predicted_ints = predicted_ints[::-1]
predicted_sentence = ints_to_sentence(predicted_ints)
sent_set.append(predicted_sentence)
# load LM to re-rank all candidates
if self.re_ranking:
sent_set = self.evaluator(sent_set, self.n)
idx = 0
if len(sent_set) > 1:
idx = randint(0, len(sent_set) - 2)
else:
idx = randint(0, len(sent_set) - 1)
# output all candidates
# all_sents = '|'.join(sent_set)
# print all_sents
sentences.append(sent_set[idx])
else:
predicted_line = predicted_batch[
0] # predicted is a batch of one line
predicted_line_clean = predicted_line[:
-1] # remove the end token
predicted_ints = map(
lambda x: x[0], predicted_line_clean
) # Flatten from [time_step, 1] to [time_step]
if FLAGS.rev_data:
predicted_ints = predicted_ints[::-1]
predicted_sentence = ints_to_sentence(predicted_ints)
sentences.append(predicted_sentence)
# remove space
for s in sentences:
s = re.sub(' ', '', s)
sentences_no_space.append(s)
return sentences_no_space
