# DONOTCHANGE (You can decide how often you want to save the model)
nsml.save(epoch)
def on_batch_end(batch, logs):
if batch % 50 == 1 or batch == batches_per_epoch:
print("Batch: {}/{}, Loss: {}".format(batch, batches_per_epoch,
logs['loss']))
callback = LambdaCallback(on_epoch_end=on_epoch_end,
on_batch_end=on_batch_end)
model.fit(
x=[dataset.input1, dataset.input2],
y=dataset.labels,
batch_size=config.batch,
epochs=config.epochs,
verbose=2,
callbacks=[callback],
validation_split=0.05,
shuffle=True,
)
# 로컬 테스트 모드일때 사용합니다
# 결과가 아래와 같이 나온다면, nsml submit을 통해서 제출할 수 있습니다.
# [(0.0, 9.045), (0.0, 5.91), ... ]
elif config.mode == 'test_local':
with open(os.path.join(DATASET_PATH, 'train/train_data'),
'rt',
encoding='utf-8') as f:
reviews = f.readlines()
res = nsml.infer(reviews)
print(res)
feed_dict={x: data, y_: labels, dropout_rate: config.dropout})
print('Batch : ', i + 1, '/', one_batch_size,
', BCE in this minibatch: ', float(loss))
avg_loss += float(loss)
print('epoch:', epoch,
' train_loss:', float(avg_loss/one_batch_size),
' accuracy:',
calculate_accuracy(sess.run(output_sigmoid,
feed_dict={x: test_data, y_: test_label, dropout_rate: 1.0}),
test_label)
)
nsml.report(summary=True, scope=locals(), epoch=epoch, epoch_total=config.epochs,
train__loss=float(avg_loss/one_batch_size), step=epoch)
# DONOTCHANGE (You can decide how often you want to save the model)
nsml.save(epoch)
# 로컬 테스트 모드일때 사용합니다
# 결과가 아래와 같이 나온다면, nsml submit을 통해서 제출할 수 있습니다.
# [(0.3, 0), (0.7, 1), ... ]
elif config.mode == 'test_local':
with open(os.path.join(DATASET_PATH, 'train/train_data'), 'rt', encoding='utf-8') as f:
queries = f.readlines()
res = nsml.infer(queries)
print(res)
with open(os.path.join(DATASET_PATH, 'train/train_label')) as f:
labels = np.array([[np.float32(x)] for x in f.readlines()])
predictions = np.array(res)[:,1]
accuracy = np.array(predictions > config.threshold).sum()/float(len(labels))
print('Accuracy: ', accuracy)
y_: labels
})
# _, right_loss = sess.run([train_step, rmse], feed_dict={x: left_data, y_: labels})
loss = float(loss)
print('Batch : ', i + 1, '/', one_batch_size,
', RMSE in this minibatch: ', loss)
avg_loss += loss
print('epoch:', epoch, ' train_loss:', avg_loss / one_batch_size)
nsml.report(summary=True,
scope=locals(),
epoch=epoch,
epoch_total=config.epochs,
train__loss=avg_loss / one_batch_size,
step=epoch)
# DONOTCHANGE (You can decide how often you want to save the model)
nsml.save(epoch)
# 로컬 테스트 모드일때 사용합니다
# 결과가 아래와 같이 나온다면, nsml submit을 통해서 제출할 수 있습니다.
# [(0.3, 0), (0.7, 1), ... ]
elif config.mode == 'test_local':
with open(os.path.join(DATASET_PATH, 'train/train_data'),
'rt',
encoding='utf-8') as f:
queries = f.readlines()
res = []
for batch in _batch_loader(queries, config.batch):
temp_res = nsml.infer(batch)
res += temp_res
print(res)
epoch_total=config.epochs,
train__loss=float(avg_loss / one_batch_size),
step=epoch,
accuracy=float(accuracy))
# DONOTCHANGE (You can decide how often you want to save the model)
nsml.save(epoch)
if not HAS_DATASET and not IS_ON_NSML:
local_save(sess, epoch)
# 로컬 테스트 모드일때 사용합니다
# 결과가 아래와 같이 나온다면, nsml submit을 통해서 제출할 수 있습니다.
# [(0.3, 0), (0.7, 1), ... ]
elif config.mode == 'test_local':
if not HAS_DATASET and not IS_ON_NSML:
local_load(sess)
with open(os.path.join(DATASET_PATH, 'train/train_data'),
'rt',
encoding='utf-8') as f:
queries = f.readlines()
res = []
for batch in _batch_loader(queries, config.batch):
batch_size = len(batch)
if USE_GPU and batch_size < config.batch:
batch += [".\t."] * (config.batch - batch_size)
temp_res = nsml.infer(batch)[:batch_size]
else:
temp_res = nsml.infer(batch)
res += temp_res
print(temp_res)
