def dev_step(sess, cnn, testList, dev_size=100):
scoreList = []
if dev_size > len(testList) / 500:
dev_size = len(testList) / 500
print("have test %d samples" % dev_size)
for i in range(dev_size):
batch_scores = []
for j in range(int(500 / FLAGS.batch_size)):
x_test_1, x_test_2, x_test_3 = insurance_qa_data_helpers.load_val_batch(
testList, vocab, i * 500 + j * FLAGS.batch_size,
FLAGS.batch_size)
feed_dict = {
cnn.input_x_1: x_test_1,
cnn.input_x_2:
x_test_2, # x_test_2 equals x_test_3 for the test case
cnn.input_x_3: x_test_3
}
predicted = sess.run(cnn.score12, feed_dict)
batch_scores.extend(predicted)
max_indexs = []
max_score = max(batch_scores)
for index, s in enumerate(batch_scores):
if s == max_score:
max_indexs.append(index)
if len(max_indexs) == 0:
scoreList.append(0)
continue
index = random.choice(max_indexs)
if int(testList[i * 500 + index].split()[0]) == 1:
scoreList.append(1)
else:
scoreList.append(0)
return sum(scoreList) * 1.0 / len(scoreList)
def dev_step(sess, cnn, testList, dev_size=100):
scoreList = []
if dev_size > len(testList) / 500:
dev_size = len(testList) / 500
print("have test %d samples" % dev_size)
for i in range(dev_size):
#１ループ内は１つのqに対する500の候補処理
batch_scores = []
for j in range(int(500 / FLAGS.batch_size)):
#x_test_1:test1のq、 x_test_2とx_test_3は同じで:test2のaかneg
x_test_1, x_test_2, x_test_3 = insurance_qa_data_helpers.load_val_batch(
testList, vocab, i * 500 + j * FLAGS.batch_size,
FLAGS.batch_size)
feed_dict = {
cnn.input_x_1: x_test_1,
cnn.input_x_2:
x_test_2, #x_test_2 equals x_test_3 for the test case
cnn.input_x_3: x_test_3
# cnn.dropout_keep_prob: 1.0
}
#predictedはリスト形式, score12はcos(q,pos)
predicted = sess.run(
[cnn.score12],
feed_dict) # add the [] means the returned value is a list
batch_scores.extend(predicted[0])
max_indexs = []
max_score = max(batch_scores)
for index, s in enumerate(batch_scores):
if s == max_score:
max_indexs.append(index) #score12が最大の候補のインデックスを探索
index = random.choice(max_indexs)
print('ラベル 質問ID 質問文 正解回答文')
print(testList[i * 500].split())
print('ラベル 質問ID 質問文 モデルの予測回答文')
print(testList[i * 500 + index].split())
print('')
if int(testList[i * 500 + index].split()[0]) == 1:
#予測した候補のラベルが１ならscoreを加算
scoreList.append(1)
else:
scoreList.append(0)
return sum(scoreList) * 1.0 / len(scoreList)
def dev_step(sess, cnn, testList, dev_size=100):
score_rr_list = []
score_pre_list = []
if dev_size > len(testList) / 500:
dev_size = len(testList) / 500
print("have test %d samples" % dev_size)
for i in range(dev_size):
#１ループ内は１つのqに対する500の候補処理
batch_scores = []
for j in range(int(500 / FLAGS.batch_size)):
#x_test_1:test1のq、 x_test_2とx_test_3は同じで:test2のaかneg
x_test_1, x_test_2, x_test_3 = insurance_qa_data_helpers.load_val_batch(
testList, vocab, i * 500 + j * FLAGS.batch_size,
FLAGS.batch_size)
feed_dict = {
cnn.input_x_1: x_test_1,
cnn.input_x_2:
x_test_2, #x_test_2 equals x_test_3 for the test case
cnn.input_x_3: x_test_3
# cnn.dropout_keep_prob: 1.0
}
#predictedはリスト形式, score12はcos(q,pos)
predicted = sess.run(
[cnn.score12],
feed_dict) # add the [] means the returned value is a list
batch_scores.extend(predicted[0])
#MRR
score_rr = rr_for_mrr(i, batch_scores, testList)
score_rr_list.append(score_rr)
#precision@k
score_at_k = precision_at(i, batch_scores, testList, 10)
score_pre_list.append(score_at_k)
mrr = sum(score_rr_list) * 1.0 / len(score_rr_list)
precision = sum(score_pre_list) * 1.0 / len(score_pre_list)
return mrr, precision