Path: blob/master/site/ko/guide/migrate/mirrored_strategy.ipynb
25118 views
Copyright 2021 The TensorFlow Authors.
단일 작업자 멀티 GPU 훈련 마이그레이션하기
이 가이드는 단일 작업자 멀티 GPU 워크플로를 TensorFlow 1에서 TensorFlow 2로 마이그레이션하는 방법을 설명합니다.
다음과 같이 한 머신에서 멀티 GPU로 동기식 훈련을 수행할 수 있습니다.
TensorFlow 1에서는
tf.distribute.MirroredStrategy
와 함께tf.estimator.Estimator
API를 사용합니다.TensorFlow 2에서는
tf.distribute.MirroredStrategy
와 함께 Keras Model.fit 또는 사용자 정의 훈련 루프를 사용할 수 있습니다. TensorFlow를 사용하는 분산 훈련 가이드에서 자세히 알아보세요.
설치하기
데모를 위해 가져오기 및 간단한 데이터세트로 시작해 보겠습니다.
TensorFlow 1: tf.estimator.Estimator로 단일 작업자 분산 훈련하기
이 예제는 단일 작업자 멀티 GPU 훈련의 TensorFlow 1 정식 워크플로를 보여줍니다. tf.estimator.Estimator
의 config
매개변수를 통해 배포 전략(tf.distribute.MirroredStrategy
)을 설정해야 합니다.
TensorFlow 2: Keras로 단일 작업자 훈련하기
TensorFlow 2로 마이그레이션하는 경우 tf.distribute.MirroredStrategy
와 함께 Keras API를 사용할 수 있습니다.
모델 빌드에 tf.keras
API를 사용하고 훈련에 Keras Model.fit
을 사용하는 경우 주요 차이점은 tf.estimator.Estimator
에 config
를 정의하는 대신 Strategy.scope
컨텍스트에서 Keras 모델, 옵티마이저 및 메트릭을 인스턴스화하는 것입니다.
사용자 정의 훈련 루프를 사용해야 하는 경우 사용자 정의 훈련 루프와 함께 tf.distribute.Strategy 사용하기 가이드를 확인하세요.
다음 단계
TensorFlow 2에서 tf.distribute.MirroredStrategy
를 사용하는 분산 훈련에 대한 자세한 내용은 다음 설명서를 확인하세요