def dev_step(sess, model, target_loss_weight):
results = defaultdict(list)
num_test = 0
num_correct = 0.0
test_batches = data_helpers.batch_iter(valid_dataset,
FLAGS.batch_size,
1,
target_loss_weight,
FLAGS.max_sequence_length,
charVocab,
FLAGS.max_word_length,
shuffle=True)
for test_batch in test_batches:
x_question, x_answer, x_question_len, x_answer_len, x_target, x_target_weight, id_pairs, x_q_char, x_q_len, x_a_char, x_a_len = test_batch
feed_dict = {
model.question: x_question,
model.answer: x_answer,
model.question_len: x_question_len,
model.answer_len: x_answer_len,
model.target: x_target,
model.target_loss_weight: x_target_weight,
model.dropout_keep_prob: 1.0,
model.q_charVec: x_q_char,
model.q_charLen: x_q_len,
model.a_charVec: x_a_char,
model.a_charLen: x_a_len
}
batch_accuracy, predicted_prob = sess.run(
[model.accuracy, model.probs], feed_dict)
num_test += len(predicted_prob)
if num_test % 1000 == 0:
print(num_test)
num_correct += len(predicted_prob) * batch_accuracy
for i, prob_score in enumerate(predicted_prob):
question_id, answer_id, label = id_pairs[i]
results[question_id].append((answer_id, label, prob_score))
#calculate top-1 precision
print('num_test_samples: {} test_accuracy: {}'.format(
num_test, num_correct / num_test))
accu, precision, recall, f1, loss = metrics.classification_metrics(results)
print('Accuracy: {}, Precision: {} Recall: {} F1: {} Loss: {}'.format(
accu, precision, recall, f1, loss))
mrr = metrics.mean_reciprocal_rank(results)
top_1_precision = metrics.top_k_precision(results, k=1)
top_2_precision = metrics.top_k_precision(results, k=2)
top_5_precision = metrics.top_k_precision(results, k=5)
top_10_precision = metrics.top_k_precision(results, k=10)
total_valid_query = metrics.get_num_valid_query(results)
print(
'MRR (mean reciprocal rank): {}\tTop-1 precision: {}\tNum_query: {}\n'.
format(mrr, top_1_precision, total_valid_query))
print('Top-2 precision: {}'.format(top_2_precision))
print('Top-5 precision: {}'.format(top_5_precision))
print('Top-10 precision: {}'.format(top_10_precision))
return mrr
def dev_step():
results = defaultdict(list)
num_test = 0
num_correct = 0.0
test_batches = data_helpers.batch_iter(test_dataset,
FLAGS.batch_size,
1,
target_loss_weight,
idf,
SEQ_LEN,
charVocab,
FLAGS.max_word_length,
shuffle=True)
for test_batch in test_batches:
x_question, x_answer, x_question_len, x_answer_len, x_target, x_target_weight, id_pairs, extra_feature, q_feature, a_feature, x_q_char, x_q_len, x_a_char, x_a_len = test_batch
feed_dict = {
esim.question: x_question,
esim.answer: x_answer,
esim.question_len: x_question_len,
esim.answer_len: x_answer_len,
esim.target: x_target,
esim.target_loss_weight: x_target_weight,
esim.dropout_keep_prob: 1.0,
esim.extra_feature: extra_feature,
esim.q_word_feature: q_feature,
esim.a_word_feature: a_feature,
esim.q_charVec: x_q_char,
esim.q_charLen: x_q_len,
esim.a_charVec: x_a_char,
esim.a_charLen: x_a_len
}
batch_accuracy, predicted_prob = sess.run(
[esim.accuracy, esim.probs], feed_dict)
num_test += len(predicted_prob)
if num_test % 1000 == 0:
print(num_test)
num_correct += len(predicted_prob) * batch_accuracy
for i, prob_score in enumerate(predicted_prob):
question_id, answer_id, label = id_pairs[i]
results[question_id].append(
(answer_id, label, prob_score))
#calculate top-1 precision
print('num_test_samples: {} test_accuracy: {}'.format(
num_test, num_correct / num_test))
accu, precision, recall, f1, loss = metrics.classification_metrics(
results)
print(
'Accuracy: {}, Precision: {} Recall: {} F1: {} Loss: {}'.
format(accu, precision, recall, f1, loss))
mvp = metrics.mean_average_precision(results)
mrr = metrics.mean_reciprocal_rank(results)
top_1_precision = metrics.top_1_precision(results)
total_valid_query = metrics.get_num_valid_query(results)
print(
'MAP (mean average precision: {}\tMRR (mean reciprocal rank): {}\tTop-1 precision: {}\tNum_query: {}'
.format(mvp, mrr, top_1_precision, total_valid_query))
return mrr
q_char_len: x_q_len,
a_char_feature: x_a_char,
a_char_len: x_a_len
}
predicted_prob = sess.run(prob, feed_dict)
num_test += len(predicted_prob)
print('num_test_sample={}'.format(num_test))
for i, prob_score in enumerate(predicted_prob):
qid, aid, label = batch_id_pairs[i]
results[qid].append((aid, label, prob_score))
accu, precision, recall, f1, loss = metrics.classification_metrics(results)
print('Accuracy: {}, Precision: {} Recall: {} F1: {} Loss: {}'.format(
accu, precision, recall, f1, loss))
mrr = metrics.mean_reciprocal_rank(results)
top_1_precision = metrics.top_k_precision(results, k=1)
top_2_precision = metrics.top_k_precision(results, k=2)
top_5_precision = metrics.top_k_precision(results, k=5)
top_10_precision = metrics.top_k_precision(results, k=10)
total_valid_query = metrics.get_num_valid_query(results)
print('MRR (mean reciprocal rank): {}\tTop-1 recall: {}\tNum_query: {}'.format(
mrr, top_1_precision, total_valid_query))
print('Top-2 recall: {}'.format(top_2_precision))
print('Top-5 recall: {}'.format(top_5_precision))
print('Top-10 recall: {}'.format(top_10_precision))
out_path = FLAGS.output_file
print("Saving evaluation to {}".format(out_path))
with open(out_path, 'w') as f:
f.write("query_id\tdocument_id\tscore\trank\trelevance\n")