텐서플로 120% 활용하기¶
목 차¶
- 왜 텐서플로인가?
- 주요 업데이트
- 구글의 빅피쳐
- 텐서플로 개발 꿀팁s
- Architecture 설계
- layers, losses, metrics
- TFRecord, Dataset
- Estimator
- Export
- OrbisAI에서는..
왜 텐서플로인가?¶
배우기 어려움 → 할 줄 알면 엄청 멋있음
다양한 언어, 다양한 platform 활용 가능 → Flexibility 매우 뛰어남(TPU사용 가능)
데이터가 많으면 많을 수록 실력 발휘
파생 플러그인, 연동 Framework 풍부
강력한 High Level API
주요 업데이트¶
open ~ 0.9
placeholder, Variable, constant, queue runners, Readers
1.0 ~ 1.2
High Level APIs
1.2 ~ current
cuda Version ups, eager mode, extention estimator, performance optimization
구글의 빅피쳐¶
"Engineering mind"가 적극적으로 반영된 Framework
Tensorflow 세계 점유율 압도적 1위 → 기술을 이끄는 것은 산업
이러한 service productivity를 높이기 위한 많은 노력의 산물 Tensorflow
연구자와 개발자 모두가 편리하게 사용 할 수 있는 High Level API를 제공함으로써 협업 능률 극대화
텐서플로 꿀팁s¶
Architecture 설계¶
In [4]:
Image('./tensorflow_presentation_images/image1.png')
Out[4]:
데이터를 서비스로 모은다면?
서비스 사용자가 늘어난다면?
데이터가 엄청나게 커진다면?
더 향상된 Performance가 요구된다면?
Pipeline, Model의 규모가 커진다면?
In [5]:
Image('./tensorflow_presentation_images/image2.png')
Out[5]:
연구자는 모델에 집중, 개발자는 서비스에 집중.¶
TF - Service 간 연동 시 소모되는 시간과 노력 최소화¶
In [ ]:
# FFNN
l1 = tf.layers.dense(x, 10, activation=tf.nn.relu)
l2 = tf.layers.dense(l1, 10, activation=tf.nn.relu)
l3 = tf.layers.dense(l2, 10, activation=tf.nn.relu)
out = tf.layers.desne(l3, 3, activation=tf.nn.softmax)
In [ ]:
# CNN
conv1 = tf.layers.conv2d(x, filters=16, kernel_size=3)
pool1 = tf.layers.max_pooling2d(conv1, pool_size=2, strides=2)
conv2 = tf.layers.conv2d(pool1, filters=32, kernel_size=[5, 4])
pool2 = tf.layers.max_pooling2d(conv2, pool_size=2, strides=2)
flat = tf.layers.flatten(pool2)
dropout_flat = tf.layers.dropout(flat, rate=0.3, training=is_training)
fc1 = tf.layers.dense(dropout_flat, 1000)
dropout_fc1 = tf.layers.dropout(fc1, rate=0.3, training=is_training)
fc2 = tf.layers.dense(dropout_fc1, 100, activation=tf.nn.relu)
dropout_fc2 = tf.layers.dropout(fc2, rate=0.3, training=is_training)
out = tf.layers.dense(dropout_fc2, 3, activation=tf.nn.softmax)
In [ ]:
## RNN
cell = tf.contrib.rnn.BasicRNNCell(128)
output, state = tf.nn.dynamic_rnn(cell, x, sequence_length=lengths, dtype=tf.float32)
In [ ]:
### loss functions
mse_loss = tf.losses.mean_squared_error(label, prediction)
abs_loss = tf.losses.absolute_difference(label, prediction)
sum_loss = tf.losses.add_loss(label, prediction)
cosd_loss = tf.losses.cosine_distance(label, prediction)
hinge_loss = tf.losses.hinge_loss(label, prediction)
In [ ]:
### metrics
tf.metrics.accuracy(label, prediction)
tf.metrics.auc(label, prediction)
tf.metrics.average_precision_at_k(label, prediction)
tf.metrics.false_negatives(label, prediction)
tf.metrics.false_negatives_at_thresholds(label, prediction)
tf.metrics.false_positives(label, prediction)
tf.metrics.false_positives_at_thresholds(label, prediction)
tf.metrics.mean(label, prediction)
tf.metrics.mean_absolute_error(label, prediction)
tf.metrics.mean_cosine_distance(label, prediction)
tf.contrib.metrics.auc_using_histogram(label, scores)
tf.contrib.metrics.auc_using_histogram(label, scores)
tf.contrib.metrics.confusion_matrix(labels, prediction)
tf.contrib.metrics.recall_at_precision(label, prediction)
한 번의 open으로 training간 reading만 수행 → file open에서 생기는 병목 최소화
data reading에 최적화된 serializing 기술로 scan 속도 매우 빠름
HDFS와 같은 대용량 File System에 적합
Dataset 자체적으로 reading시 multi-processing 수행
pipeline 구성이 매우 간결해짐
In [ ]:
### TFRecord Transform
with tf.python_io.TFRecordWriter('/path/to/write/file.tfrecord') as write:
example = tf.train.Example()
example.features.feature['key1'].float_list.value.extend(float_list)
example.features.feature['key2'].int64_list.value.extend(int_list)
example.features.feature['key3'].float_list.value.append(float_value)
example.features.feature['key4'].int64_list.value.append(int_value)
serialized = example.SerializeToString()
writer.write(serialized)
In [ ]:
### TFRecord parser
def parser(serialized_example):
feature = {
'key1': tf.FixedLenFeature([10], tf.float32),
'key2': tf.VarLenFeature(tf.int64),
'key3': tf.FixedLenFeature([1], tf.float32),
'key4': tf.FixedLenFeature([1], tf.int64),
}
parsed_feature = tf.parse_single_example(serialized_example, feature)
feature1 = parsed_feature['key1']
feature2 = parsed_feature['key2']
feature3 = parsed_feature['key3']
label = parsed_feature['key4']
return feature1, feature2, feature3, label
In [ ]:
### Reading TFRecord
dataset = tf.data.TFRecordDataset('/path/from/read/file.tfrecord')
dataset = dataset.map(parser)
dataset = dataset.batch(32).shuffle(50000).repeat(100)
itr = dataset.make_one_shot_iterator()
feature1, feature2, feature3, label = itr.get_next()
끝나지 않은 코딩에(Session...) 지친 당신¶
Estimator로 일어나라아아ㅏㅏㅏ¶
checkpoint 자동 저장, 저장된 모델 자동 restore
자체 export 기능
train, evaluate, prediction 간편하게 관리
pipeline, model, runner을 블럭 조립하듯이 모듈로 사용 가능
코드 줄 대폭 감소, 깔끔한 추상화로 가독성 증가
In [ ]:
est = tf.estimator.Estimator(model_fn, model_dir)
est.train(train_input_fn)
est.evaluate(eval_input_fn)
est.predict(pred_input_fn)
In [ ]:
def input_fn():
dataset = tf.data.TFRecordDataset('/path/from/read/file.tfrecord')
dataset = dataset.map(parser)
dataset = dataset.batch(32).shuffle(50000).repeat(100)
itr = dataset.make_one_shot_iterator()
feature1, feature2, feature3, label = itr.get_next()
return {'feature1': feature1,
'feature2': feature2,
'feature3': feature3,
}, label
In [ ]:
def model_fn(features, label, mode):
### model code here ###
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.train.get_global_step()
loss = ...
train_op = ...minimize(loss, global_step)
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
train_op=train_op,
loss=loss)
elif mode == tf.estimator.ModeKeys.EVAL:
loss = ...
metirc = ...
eval_metric_ops = {'metric': metric}
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
elif mode == tf.estimator.ModeKeys.PRED:
prediction = {'output1': output1, 'output2': output2}
estimator_spec = tf.estimator.EstimatorSpec(mode=mode,
predictions=predictions)
else:
raise Exception('invalid estimator')
return estimator_spec
Serving 가즈아아ㅏㅏ¶
C++ 기반 서버
grpc 프로토콜 사용
server down 없이 model update 가능
In [ ]:
### Serving receiver function
def _serving_input_receiver_fn():
feature = {
'key1': tf.FixedLenFeature([10], tf.float32),
'key2': tf.VarLenFeature(tf.int64),
'key3': tf.FixedLenFeature([1], tf.float32),
}
serialized_example = tf.placeholder(tf.string)
receiver_tensors = {'example': serialized_example}
features = tf.parse_example(serialized_example, feature)
return tf.estimator.export.ServingInputReceiver(features, receiver_tensors)
In [ ]:
### export output
elif PRED:
prediction = {'output1': output1, 'output2': output2}
estimator_spec = tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
export_outputs={'response': tf.estimator.export.PredictOutput(predictions)})
In [ ]:
est = tf.estimator.Estimator(model_fn, model_dir)
est.export_savedmodel(export_dir_base='/path/to/export',
serving_input_receiver_fn=serving_input_receiver_fn)
In [ ]:
$ tensorflow_model_server --port=port-numbers --model_name=model-name-you-wanted --model_base_path=/absolute/path/of/your/saved/model
tensorflow model server api installation guide
https://www.tensorflow.org/serving/setup
client example
https://github.com/.../tensor.../example/inception_client.py
OrbisAI에서는...¶
In [6]:
Image('./tensorflow_presentation_images/image3.png')
Out[6]:
In [8]:
Image('./tensorflow_presentation_images/image4.png')
Out[8]:
감사합니다¶
텐서플로 120% 활용하기