underlying storage in the cloud (AWS EBS, AWS S3, AWS Glacier, aliyun EBS, aliyun OSS) or on-prem (baremetal, HDFS, OSS) and turns it into container-native storageDatabase • Database services orchestrated ) • CurveFS can manage different storages (HDFS, OSS, EBS) below • Apps access data by POSIX interface • Infrequent data is moved to OSS, and frequent data is moved to high speed storage transparentlyData

tf.Adamoptimizer(learning_rate=...) # Choose a 1oss Scale manager which decides how to pick the right loss scale # throughout the training process. 1oss_scale_manger = tf.contrib.mixed_precision.Fixed loss_scale_optimizer = LossScaleOptimizer(opt，1oss_scale_manager) # Call minimize() on the loss scale optimizer. train_op = loss_scale_optimizer.minimize(1oss) Loss Scaling in PAI-TF Loss Scaling the loss