def _predict_pos_tags(self, session, model, words, data_path):
'''
Define prediction function of POS Tagging
return tuples [(word, tag)]
'''
word_data = pos_reader.sentence_to_word_ids(data_path, words)
tag_data = [0] * len(word_data)
state = session.run(model.initial_state)
predict_id = []
for step, (x, y) in enumerate(
pos_reader.iterator(word_data, tag_data, model.batch_size,
model.num_steps)):
#print ("Current Step" + str(step))
fetches = [model.cost, model.final_state, model.logits]
feed_dict = {}
feed_dict[model.input_data] = x
feed_dict[model.targets] = y
for i, (c, h) in enumerate(model.initial_state):
feed_dict[c] = state[i].c
feed_dict[h] = state[i].h
_, _, logits = session.run(fetches, feed_dict)
predict_id.append(int(np.argmax(logits)))
#print (logits)
predict_tag = pos_reader.word_ids_to_sentence(data_path, predict_id)
return zip(words, predict_tag)
def predict(words):
word_data = reader.sentence_to_word_ids(FLAGS.pos_data_path, words)
tag_data = [0] * len(word_data)
config = get_config(FLAGS.pos_lang)
config.batch_size = 1
# config.num_steps, num_steps is T, should not be set to 1, CRF decode whe whole sequence at one time
# intialize model
ckpt_path = os.path.join(FLAGS.pos_train_dir, "pos_bilstm_crf.ckpt")
with tf.Graph().as_default(), tf.Session() as session:
with tf.variable_scope(
"pos_var_scope"): #Need to Change in Pos_Tagger Save Function
model = POSTagger(is_training=False,
config=config) # save object after is_training
if len(glob.glob(ckpt_path + '.data*')
) > 0: # file exist with pattern: 'pos.ckpt.data*'
print("Loading model parameters from %s" % ckpt_path)
all_vars = tf.global_variables()
model_vars = [
k for k in all_vars if k.name.startswith("pos_var_scope")
]
tf.train.Saver(model_vars).restore(session, ckpt_path)
else:
print("Model not found, created with fresh parameters.")
session.run(tf.global_variables_initializer())
# Make Prediciton
word_data = np.array(word_data, dtype=np.int32)
sequence_len = len(word_data)
x = np.zeros([1, config.num_steps], dtype=np.int32)
for i in range(sequence_len): # pad word_data len to num_steps
x[0, i] = word_data[i]
y = np.zeros([1, config.num_steps], dtype=np.int32)
fetches = [model.cost, model.logits, model.transition_params]
feed_dict = {}
feed_dict[model.input_data] = x
feed_dict[model.targets] = y
cost, logits, transition_params = session.run(fetches, feed_dict)
print(logits[0])
print(logits)
# print (transition_params)
print(transition_params[0, :])
# Remove padding from the scores and tag sequence.
tf_unary_scores = logits[
0, :, :] # One Example [batch_size, num_steps, target_num]
print("unary_score %d" % len(tf_unary_scores))
tf_unary_scores = tf_unary_scores[:sequence_len]
print("unary_score %d" % len(tf_unary_scores))
# transition_params [Aij] tag transition probability
viterbi_sequence, _ = tf.contrib.crf.viterbi_decode(
tf_unary_scores, transition_params)
print("optimal sequence found by viterbi algorithm")
print(viterbi_sequence)
predict_tag = reader.word_ids_to_sentence(FLAGS.pos_data_path,
viterbi_sequence)
return predict_tag
