docker pull tensorflow/tensorflow:1.13.1-py3-jupyter docker pull tensorflow/serving:1.13.0

docker run --name my_tf -it -p 8881-8889:8881-8889 -v C:\notebooks\:/notebooks/ tensorflow/tensorflow:1.13.1-py3-jupyter bash

cd /notebooks && jupyter notebook --allow-root --ip=0.0.0.0 --port=8881 &

# 필요한 모듈을 추가로 설치한다 !pip install sklearn # 사용할 모듈을 불러온다 from os import path, listdir from shutil import copytree from sklearn import datasets from sklearn.model_selection import train_test_split import tensorflow as tf # 텐서플로 모듈을 불러온다 print('tensorflow version:', tf.__version__) from tensorflow import estimator iris = datasets.load_iris() # Iris 데이터 읽어온다 x = iris.data y = iris.target # 트레이닝셋과 검증셋으로 분리한다 x_train, x_eval, y_train, y_eval = train_test_split(x, y, test_size=0.2) model_dir = './iris/' # 모델을 저장할 디렉토리 model_name = 'iris_dnn' # 모델명 num_epochs = 20 # 에폭(입력 데이터를 몇회 순환할지) train_input_fn = estimator.inputs.numpy_input_fn(x={'x':x_train}, y=y_train, num_epochs=num_epochs, shuffle=True) train_spec = estimator.TrainSpec(input_fn=train_input_fn) eval_input_fn = estimator.inputs.numpy_input_fn(x={'x':x_eval}, y=y_eval, num_epochs=1, shuffle=False) # 서빙의 입력 메타를 정의합니다 def serving_input_receiver_fn(): receiver_tensors = { 'sepal_length': tf.placeholder(tf.float32, [None, 1]), 'sepal_width' : tf.placeholder(tf.float32, [None, 1]), 'petal_length': tf.placeholder(tf.float32, [None, 1]), 'petal_width' : tf.placeholder(tf.float32, [None, 1]), } features = { 'x': tf.concat([ receiver_tensors['sepal_length'], receiver_tensors['sepal_width'], receiver_tensors['petal_length'], receiver_tensors['petal_width'] ], axis=1) } return tf.estimator.export.ServingInputReceiver(receiver_tensors=receiver_tensors, features=features) # 최근 모델을 저장할 Exporter latest_exporter = estimator.LatestExporter(name='latest_exporter', serving_input_receiver_fn=serving_input_receiver_fn) # 가장 좋은 모델을 저장할 Exporter best_exporter = estimator.BestExporter(name='best_exporter', serving_input_receiver_fn=serving_input_receiver_fn) exporters = [latest_exporter, best_exporter] eval_spec = estimator.EvalSpec(input_fn=eval_input_fn, throttle_secs=10, exporters=exporters) classifier = estimator.DNNClassifier(config=estimator.RunConfig(model_dir=model_dir), feature_columns=[tf.feature_column.numeric_column('x', shape=[4])], hidden_units=[20], n_classes=3, model_dir=model_dir) tf.estimator.train_and_evaluate(classifier, train_spec=train_spec, eval_spec=eval_spec) def get_abs_directory_list(in_path): """입력된 경로의 (절대경로)디렉토리 목록을 가져옵니다""" out_paths = [] out_ids = [] if path.exists(in_path) and len(listdir(in_path)) >=1: for id in listdir(in_path): abs_dir = path.join(in_path, id) if path.isdir(abs_dir): out_paths.append(abs_dir) out_ids.append(id) return out_paths, out_ids # 가장 좋은 모델이 저장된 디렉토리 best_exporter_path = path.join(model_dir, 'export', 'best_exporter') src_paths, src_ids = get_abs_directory_list(best_exporter_path) # 서빙되고 있는 모델이 저장된 디렉토리 serving_exporter_path = path.join(model_dir, 'export', 'serving_exporter', model_name) _, des_ids = get_abs_directory_list(serving_exporter_path) for idx, src_path in enumerate(src_paths): # 순회 if src_ids[idx] not in des_ids: # 신규 모델이라면 copytree(src_path, path.join(serving_exporter_path, src_ids[idx])) # 복사한다 print(str(src_ids[idx]) + ' copy!')

docker run --rm --name serving_1 -v C:\notebooks\iris\export\serving_exporter:/models/ -p 8501:8501 -e MODEL_NAME=iris_dnn -e MODEL_BASE_PATH=/models tensorflow/serving:1.13.0 docker run --rm --name serving_2 -v C:\notebooks\iris\export\serving_exporter:/models/ -p 8502:8501 -e MODEL_NAME=iris_dnn -e MODEL_BASE_PATH=/models tensorflow/serving:1.13.0 docker run --rm --name serving_3 -v C:\notebooks\iris\export\serving_exporter:/models/ -p 8503:8501 -e MODEL_NAME=iris_dnn -e MODEL_BASE_PATH=/models tensorflow/serving:1.13.0

POST http://host:port/v1/models/${MODEL_NAME}[/versions/${MODEL_VERSION}]:(predict|classify|regress) # /versions/${MODEL_VERSION}는 생략가능하며 생략 시 가장 최신 모델이 실행됩니다 요청예시) POST http://host:port/v1/models/MY_MODEL:predict

{"signature_name": <string>, "instances": <value>|<(nested)list>|<list-of-objects>} 인스턴스예시) json과 유사함 "instances": [ { "f1": [1, 2, 3, 4, 5], "f2": [[1, 2], [3, 4]] }, { "f1": [3, 4, 1, 2, 5]], "f2": [[4, 5], [6, 8]] } ]

curl -X POST http://127.0.0.1:8501/v1/models/iris_dnn:predict -d '{"signature_name":"predict","instances":[{"sepal_length":[5.0],"sepal_width":[3.3],"petal_length":[1.4],"petal_width":[0.2]}]}' curl -X POST http://127.0.0.1:8501/v1/models/iris_dnn:predict -d '{"signature_name":"predict","instances":[{"sepal_length":[5.7],"sepal_width":[2.8],"petal_length":[4.1],"petal_width":[1.3]}]}' curl -X POST http://127.0.0.1:8501/v1/models/iris_dnn:predict -d '{"signature_name":"predict","instances":[{"sepal_length":[5.9],"sepal_width":[3.0],"petal_length":[5.1],"petal_width":[1.8]}]}'

curl -X POST http://127.0.0.1:8501/v1/models/iris_dnn:predict -d "{""signature_name"":""predict"",""instances"":[{""sepal_length"":[5.0],""sepal_width"":[3.3],""petal_length"":[1.4],""petal_width"":[0.2]}]}" curl -X POST http://127.0.0.1:8501/v1/models/iris_dnn:predict -d "{""signature_name"":""predict"",""instances"":[{""sepal_length"":[5.7],""sepal_width"":[2.8],""petal_length"":[4.1],""petal_width"":[1.3]}]}" curl -X POST http://127.0.0.1:8501/v1/models/iris_dnn:predict -d "{""signature_name"":""predict"",""instances"":[{""sepal_length"":[5.9],""sepal_width"":[3.0],""petal_length"":[5.1],""petal_width"":[1.8]}]}"