def training_result(num=0): # 학습 결과 확인
# 이 모델은 입력값과 출력값 데이터로 [영어단어, 한글단어] 사용하지만,
# 예측시에는 한글단어를 알지 못하므로, 디코더의 입출력값을 의미 없는 값인 P 값으로 채운다.
# ['word', 'PPPP']
start = num
end = num + 1
encoder_inputs, decoder_inputs, targets_vector, target_weights = tool.make_batch(
encoderinputs[start:end], decoderinputs[start:end],
targets_[start:end], targetweights[start:end])
encoder_vector = []
decoder_vector = []
temp_encoder = []
temp_decoder = []
for j in range(encoder_size):
temp_word = ix_to_word[encoder_inputs[j][0]][:3]
temp_encoder.append(word_to_vector[temp_word])
for j in range(decoder_size):
temp_word = ix_to_word[decoder_inputs[j][0]][:3]
temp_decoder.append(word_to_vector[temp_word])
encoder_vector.append(np.array(temp_encoder))
decoder_vector.append(np.array(temp_decoder))
targets_vector = np.transpose(targets_vector)
# 결과가 [batch size, time step, input] 으로 나오기 때문에,
# 2번째 차원인 input 차원을 argmax 로 취해 가장 확률이 높은 글자를 예측 값으로 만든다.
prediction = tf.argmax(model, 2)
result = sess.run(prediction,
feed_dict={
enc_input: encoder_vector,
dec_input: decoder_vector,
targets: targets_vector
})
# 결과 값인 숫자의 인덱스에 해당하는 글자를 가져와 글자 배열을 만든다.
# print(result[0])
result_target = ""
predict_target = ""
training_resultd = ""
for target_index in targets_vector[0]:
result_target += ix_to_word[target_index]
result_target += " "
for result_index in result[0]:
predict_target += ix_to_word[result_index]
predict_target += " "
training_resultd = (str(num) + "\ntarget : " + result_target +
"\npredict : " + predict_target)
# 출력의 끝을 의미하는 'E' 이후의 글자들을 제거하고 문자열로 만든다.
# end = decoded.index('E')
return training_resultd
start = 0
end = batch_size
index = 0
while current_step < 10000001:
if end > len(title):
start = 0
end = batch_size
if index > len(test_title):
index = 0
# Get a batch and make a step
start_time = time.time()
encoder_inputs, decoder_inputs, targets, target_weights = tool.make_batch(encoderinputs[start:end],
decoderinputs[start:end],
targets_[start:end],
targetweights[start:end])
if current_step % steps_per_checkpoint == 0:
for i in range(decoder_size - 2):
decoder_inputs[i + 1] = np.array([word_to_ix['<PAD>']] * batch_size)
output_logits = model.step(sess, encoder_inputs, decoder_inputs, targets, target_weights, True)
predict = [np.argmax(logit, axis=1)[0] for logit in output_logits]
predict = ' '.join(ix_to_word[ix] for ix in predict)
real = [word[0] for word in targets]
real = ' '.join(ix_to_word[ix] for ix in real)
print('\n----\n step : %s \n time : %s \n LOSS : %s \n 예측 : %s \n 손질한 정답 : %s \n 정답 : %s \n----' %
(current_step, step_time, loss, predict, real, title[start]))
loss, step_time = 0.0, 0.0
if (current_step) % 100 == 0:
_encoder_inputs, _decoder_inputs, _targets, _target_weights = tool.make_batch(
forward_only=forward_only)
sess.run(tf.global_variables_initializer())
step_time, loss = 0.0, 0.0
current_step = 0
start = 0
end = batch_size
while current_step < 10000001:
if end > len(title):
start = 0
end = batch_size
# Get a batch and make a step
start_time = time.time()
encoder_inputs, decoder_inputs, targets, target_weights = tool.make_batch(encoderinputs[start:end],
decoderinputs[start:end],
targets_[start:end],
targetweights[start:end])
if current_step % steps_per_checkpoint == 0:
for i in range(decoder_size - 2):
decoder_inputs[i + 1] = np.array([word_to_ix['<PAD>']] * batch_size)
output_logits = model.step(sess, encoder_inputs, decoder_inputs, targets, target_weights, True)
predict = [np.argmax(logit, axis=1)[0] for logit in output_logits]
predict = ' '.join(ix_to_word[ix][0] for ix in predict)
real = [word[0] for word in targets]
real = ' '.join(ix_to_word[ix][0] for ix in real)
print('\n----\n step : %s \n time : %s \n LOSS : %s \n 예측 : %s \n 손질한 정답 : %s \n 정답 : %s \n----' %
(current_step, step_time, loss, predict, real, title[start]))
loss, step_time = 0.0, 0.0
step_loss = model.step(sess, encoder_inputs, decoder_inputs, targets, target_weights, False)
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=model,
labels=targets)) # sparse_softmax_cross_entropy_with_logits
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(
cost) # AdamOptimizer 최적화 방법
#########
# 신경망 모델 학습
######
sess = tf.Session()
sess.run(tf.global_variables_initializer())
start = 0
end = batch_size
encoder_inputs, decoder_inputs, targets_vector, target_weights = tool.make_batch(
encoderinputs[start:end], decoderinputs[start:end], targets_[start:end],
targetweights[start:end]) # 인코더 데이터, 디코터 데이터, 타겟 데이터를 가져옵니다.
encoder_vector = []
decoder_vector = [] # 위 데이터는 transpose해서 사용해주어야합니다.
for i in range(batch_size): # 임베딩 해준거
temp_encoder = []
temp_decoder = []
for j in range(encoder_size):
temp_word = ix_to_word[encoder_inputs[j][i]][:3]
temp_encoder.append(word_to_vector[temp_word])
for j in range(decoder_size):
temp_word = ix_to_word[decoder_inputs[j][i]][:3]
# if temp_word == "<PA" : # PAD 갯수 줄이기
# if j==5:
# break
decoderinputs_test_tmp = []
targets__test_tmp = []
targetweights_test_tmp = []
test_all = 0
test_correct = 0
for k in range(0, int(test_size * (1 / batch_size))):
test_all += batch_size
test_start = k * batch_size
test_end = k * batch_size + batch_size
encoderinputs_test_tmp = encoderinputs_test[test_start:test_end]
decoderinputs_test_tmp = decoderinputs_test[test_start:test_end]
targets__test_tmp = targets__test[test_start:test_end]
targetweights_test_tmp = targetweights_test[test_start:test_end]
title_test_tmp = title_test[test_start:test_end]
encoder_inputs, decoder_inputs, targets, target_weights = tool.make_batch(
encoderinputs_test_tmp, decoderinputs_test_tmp, targets__test_tmp,
targetweights_test_tmp)
for i in range(decoder_size - 2):
decoder_inputs[i + 1] = np.array([word_to_ix['<PAD>']] * batch_size)
output_logits = model.step(sess, encoder_inputs, decoder_inputs, targets,
target_weights, True)
for i in range(0, len(np.argmax(output_logits[0], axis=1))):
# print(ix_to_word[np.argmax(output_logits[0], axis=1)[i]] + ' ' + title_tmp[i])
if title_test_tmp[i] == "직장생활":
if ix_to_word[np.argmax(output_logits[0], axis=1)[i]] == "직장생":
test_correct += 1
else:
if ix_to_word[np.argmax(output_logits[0],
axis=1)[i]] == title_test_tmp[i]:
test_correct += 1
def train(batch_size=2, epoch=100):
model = Seq2Seq(vocab_size)
with tf.Session() as sess:
# TODO: 세션을 로드하고 로그를 위한 summary 저장등의 로직을 Seq2Seq 모델로 넣을 필요가 있음
ckpt = tf.train.get_checkpoint_state("./model2")
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print("다음 파일에서 모델을 읽는 중 입니다..", ckpt.model_checkpoint_path)
model.saver.restore(sess, ckpt.model_checkpoint_path)
else:
print("새로운 모델을 생성하는 중 입니다.")
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter("./logs", sess.graph)
step_time, loss = 0.0, 0.0
current_step = 0
start = 0
end = batch_size
encoder_vector = []
decoder_vector = []
while current_step < 10000001:
if end > len(title):
start = 0
end = batch_size
# Get a batch and make a step
start_time = time.time()
encoder_inputs, decoder_inputs, targets, target_weights = tool.make_batch(
encoderinputs[start:end], decoderinputs[start:end],
targets_[start:end], targetweights[start:end])
for batch_size_ in range(batch_size): # 임베딩 해준거
temp_encoder = []
temp_decoder = []
for j in range(encoder_size):
temp_word = ix_to_word[encoder_inputs[j][batch_size_]][:3]
temp_encoder.append(word_to_vector[temp_word])
for j in range(decoder_size):
temp_word = ix_to_word[decoder_inputs[j][batch_size_]][:3]
temp_decoder.append(word_to_vector[temp_word])
encoder_vector.append(np.array(temp_encoder))
decoder_vector.append(np.array(temp_decoder))
targets_vector = np.transpose(targets)
# temp_word = ix_to_word[decoder_inputs[j][i]][:3]
# temp_decoder.append(word_to_vector[temp_word])
# encoder_vector.append(np.array(temp_encoder))
# decoder_vector.append(np.array(temp_decoder))
# for i in range(encoder_size):
#
# temp_encoder = []
# for j in range(batch_size):
# temp_word = ix_to_word[encoder_inputs[i][j]][:3]
# temp_encoder.append(word_to_vector[temp_word])
# encoder_vector.append(np.array(temp_encoder))
#
# for i in range(decoder_size):
# temp_decoder = []
# for j in range(batch_size):
# temp_word = ix_to_word[decoder_inputs[i][j]][:3]
# temp_decoder.append(word_to_vector[temp_word])
# decoder_vector.append(np.array(temp_decoder))
# for step in range(total_batch * epoch):
# enc_input, dec_input, targets = dialog.next_batch(batch_size)
_, loss = model.train(sess, encoder_vector, decoder_vector,
targets_vector)
if (step + 1) % 100 == 0:
model.write_logs(sess, writer, encoder_inputs, decoder_inputs,
targets_vector)
print('Step:', '%06d' % model.global_step.eval(), 'cost =',
'{:.6f}'.format(loss))
checkpoint_path = os.path.join("./model2", "news.ckpt")
model.saver.save(sess, checkpoint_path, global_step=model.global_step)
print('최적화 완료!')
