Method and apparatus for training deep learning model

Abstract

This disclosure embodiment disclosed methods and devices for training deep learning models.A specific implementation method of this method includes: the model description information and configuration information of the deep learning model; the model description information is divided into at least two sections according to the cut points variable in the configuration information, and the model description information is loaded to the corresponding resources to the corresponding resourcesRun; enter a batch of training samples to enter the first section of the model to describe the resources corresponding to the information and start training and use the context information obtained as the input of resources corresponding to the resources corresponding to the information;Corresponding resource operation results; if you meet the training conditions of training, the deep learning model of the output training is completed; otherwise, continue to obtain the next batch of training samples to perform the above training steps until the training complete conditions are met.This implementation method realizes the free mix of heterogeneous equipment, gives full play to the computing capabilities of different computing devices, and improves training speed.

Description

technical field [0001] Embodiments of the present disclosure relate to the field of computer technology, and in particular to methods and devices for training deep learning models. Background technique [0002] With the development of current deep learning models towards deeper levels, wider representations, and more complex structures, GPUs (Graphics Processing Units) with high-efficiency computing power have become indispensable computing resources in this field. Common parallel schemes are divided into two types: model parallelism and data parallelism. [0003] The technical solution of model parallelism distributes all model parameters to different devices to perform calculations, and each device maintains part of the parameters. Since the calculation of different devices depends on the computing context of the previous device, in order to improve the utilization of computing devices (such as GPU), the model parallel pipeline splits a large batch of data into multiple s...

AI Technical Summary

Problems solved by technology

However, this solution does not make full use of heterogeneous devices, that is, all data streams can only run on a single type of device, let alone support the ratio of computing resources of different devices

Images