def predict(sess, cnn, test, alphabet, batch_size, q_len, a_len, step, type):
scores = []
d = get_overlap_dict(test, alphabet, q_len, a_len)
for data in batch_gen_with_single(test,
alphabet,
batch_size,
q_len,
a_len,
overlap_dict=d):
feed_dict = {
cnn.question: data[0],
cnn.answer: data[1],
cnn.answer_negative: data[1],
cnn.q_pos_overlap: data[2],
cnn.q_neg_overlap: data[2],
cnn.a_pos_overlap: data[3],
cnn.a_neg_overlap: data[3],
cnn.q_position: data[4],
cnn.a_pos_position: data[5],
cnn.a_neg_position: data[5]
}
score = sess.run(cnn.score12, feed_dict)
scores.extend(score)
with open(data_file + '_' + type + "_score_%d.txt" % step, 'w') as ff:
string_tmp = '\n'.join([str(i) for i in scores])
ff.write(string_tmp)
return np.array(scores[:len(test)])
def predict(sess, cnn, test, alphabet, batch_size, q_len, a_len):
scores = []
d = get_overlap_dict(test, alphabet, q_len, a_len)
for data in batch_gen_with_single(test,
alphabet,
batch_size,
q_len,
a_len,
overlap_dict=d):
feed_dict = {
cnn.question: data[0],
cnn.answer: data[1],
cnn.answer_negative: data[1],
cnn.q_pos_overlap: data[2],
cnn.q_neg_overlap: data[2],
cnn.a_pos_overlap: data[3],
cnn.a_neg_overlap: data[3],
cnn.q_position: data[4],
cnn.a_pos_position: data[5],
cnn.a_neg_position: data[5]
}
score = sess.run(cnn.score12, feed_dict)
# print len(score)
# if batch_size == 20:
# attention.extend((q,a))
scores.extend(score)
pickle.dump(attention, open('attention.file', 'w'))
return np.array(scores[:len(test)])
def predict(sess, cnn, dev, alphabet, batch_size, q_len):
scores = []
for data in batch_gen_with_single(dev, alphabet, batch_size, q_len):
feed_dict = {
cnn.question: data[0],
cnn.q_position: data[1],
cnn.dropout_keep_prob: 1.0
}
score = sess.run(cnn.scores, feed_dict)
scores.extend(score)
return np.array(scores[:len(dev)])
def predict(sess, cnn, test, alphabet, batch_size, q_len, a_len):
scores = []
d = get_overlap_dict(test, alphabet, q_len, a_len)
for data in batch_gen_with_single(test,
alphabet,
batch_size,
q_len,
a_len,
overlap_dict=d):
feed_dict = {
cnn.question: data[0],
cnn.answer: data[1],
cnn.q_overlap: data[2],
cnn.a_overlap: data[3],
cnn.q_position: data[4],
cnn.a_position: data[5]
}
score = sess.run(cnn.scores, feed_dict)
scores.extend(score)
return np.array(scores[:len(test)])
q_max_sent_length = max(
map(lambda x: len(x), train['question'].str.split()))
alphabet, embeddings = prepare([train, test, dev],
max_sent_length=q_max_sent_length,
dim=FLAGS.embedding_dim,
is_embedding_needed=True,
fresh=True)
with tf.Session() as sess:
saver.restore(sess, ckpt.model_checkpoint_path)
graph = tf.get_default_graph()
scores = []
question = graph.get_operation_by_name('input_question').outputs[0]
q_position = graph.get_operation_by_name('q_position').outputs[0]
dropout_keep_prob = graph.get_operation_by_name(
'dropout_keep_prob').outputs[0]
for data in batch_gen_with_single(test, alphabet, FLAGS.batch_size,
q_max_sent_length):
feed_dict = {
question.name: data[0],
q_position.name: data[1],
dropout_keep_prob.name: 1.0
}
score = sess.run("output/scores:0", feed_dict)
scores.extend(score)
scores = np.array(scores[:len(test)])
predicted_label = np.argmax(scores, 1)
acc_test = acc_train = accuracy_score(predicted_label,
test['flag'])
print("test epoch:acc {}".format(acc_test))
else:
for i in range(1, FLAGS.n_fold + 1):
print("{} cross validation ".format(i))