TensorFlow 2.0 - tf.distribute 分布式训练

学习于:简单粗暴 TensorFlow 2

1. 单机多卡 MirroredStrategy

# 分布式训练
import tensorflow as tf
import tensorflow_datasets as tfds

# 1 单机多卡 MirroredStrategy

strategy = tf.distribute.MirroredStrategy()
# 指定设备
strategy = tf.distribute.MirroredStrategy(devices=['/gpu:0'])
# ------------------------------------------------
num_epochs = 5
batch_size_per_replica = 64
learning_rate = 1e-4

# 定义策略
strategy = tf.distribute.MirroredStrategy()

print("设备数量:{}".format(strategy.num_replicas_in_sync))
batch_size = batch_size_per_replica * strategy.num_replicas_in_sync


def resize(img, label):  # 处理图片
    img = tf.image.resize(img, [224, 224]) / 255.0
    return img, label


# 载入猫狗分类数据集
dataset = tfds.load("cats_vs_dogs", split=tfds.Split.TRAIN, as_supervised=True)
dataset = dataset.map(resize).shuffle(1024).batch(batch_size)

# 使用策略
with strategy.scope():
	# 模型构建代码放入 with 
    model = tf.keras.applications.MobileNetV2(weights=None, classes=2)
    model.compile(
        optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
        loss=tf.keras.losses.SparseCategoricalCrossentropy(),
        metrics=[tf.keras.metrics.sparse_categorical_accuracy]
    )

model.fit(dataset, epochs=num_epochs)

2. 多机训练 MultiWorkerMirroredStrategy

  • 相比上面,多了以下配置
  • 'task': {'type': 'worker', 'index': 0} 每台机器 index 不一样
num_workers = 2
os.environ['TF_CONFIG'] = json.dumps({
    'cluster': {
        'worker': ["localhost:20000", "localhost:20001"]
    },
    'task': {'type': 'worker', 'index': 0} # 每台机器的 index 不同
})

strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()
batch_size = batch_size_per_replica * num_workers

3. TPU 张量处理单元

可以在 Colab 上运行
在这里插入图片描述

tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
tf.config.experimental_connect_to_cluster(tpu)
tf.tpu.experimental.initialize_tpu_system(tpu)
strategy = tf.distribute.experimental.TPUStrategy(tpu)
已标记关键词 清除标记
相关推荐
©️2020 CSDN 皮肤主题: Age of Ai 设计师:meimeiellie 返回首页