Distributed Training Modules for Deep Learning Bottlenecks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The growing scale of data and models in deep learning poses a significant burden for standalone training, necessitating distributed training modes to efficiently handle large datasets and models.

Innovation Solution

The proposed solution involves a decentralized distributed training approach with independent distributed reader, trainer, and parameter server modules, where the reader acquires and loads training samples, the trainer performs model training to obtain gradient information, and the parameter server updates the model parameters, improving training speed through modular and cooperative operation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If stand-alone training is used for deep learning models, then the training process is simple to implement, but it cannot handle large-scale data and models effectively

Engineering Contradiction:
Improvecapability to handle large-scale data and modelsVSAvoidtraining system structure
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The training system is segmented into three independent functional modules: distributed reader, distributed trainer, and distributed parameter server. Each module operates independently with specific responsibilities - the reader acquires and loads training samples, the trainer performs model training and computes gradient information, and the parameter server updates model parameters. This segmentation enables the system to handle large-scale data and models while maintaining clear functional boundaries and reducing inter-module dependencies.

Inventive Principle:
Principle #1Segmentation

2Productivity

If distributed training mode is adopted to handle large-scale data and models, then training capability is improved, but module independence and coordination complexity increase

Engineering Contradiction:
Improvemodel training speedVSAvoidmodule coordination complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

Each distributed training module is designed to be self-sufficient with well-defined interfaces. The distributed reader independently acquires and loads training samples without requiring coordination for data processing. The distributed trainer independently performs model training and gradient computation. The distributed parameter server independently updates model parameters based on received gradient information. This self-service design reduces coordination overhead and simplifies module interaction while maintaining high training productivity.

Inventive Principle:
Principle #25Self-service

3Productivity

If centralized distributed training mode is used, then data management is simplified, but training efficiency and scalability are limited

Engineering Contradiction:
Improvetraining efficiencyVSAvoiddata management structure
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system segments data management responsibilities across independent modules. The distributed reader module handles data acquisition and loading, the distributed trainer module handles training data processing and gradient computation, and the distributed parameter server handles parameter storage and updates. This segmentation of data management functions enables parallel processing and improves training efficiency while distributing data management complexity across specialized modules rather than centralizing it.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4016398A1Apparatus and method for distributed training model, and computer program product
Publication Date: 2022.06.22 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP4016398A1 patent drawingFigure 1
  • EP4016398A1 patent drawingFigure 2
  • EP4016398A1 patent drawingFigure 3

AI summary

The present disclosure discloses an apparatus and method for distributed training a model, an electronic device, a computer readable storage medium, and a computer program product, and relates to the field of deep learning technology. An embodiment of the apparatus includes: a distributed reader, a distributed trainer and a distributed parameter server that are mutually independent. A reader in the distributed reader is configured to acquire a training sample, and load the acquired training sample to a corresponding trainer in the distributed trainer; the trainer in the distributed trainer is configured to perform model training based on the loaded training sample to obtain gradient information; and a parameter server in the distributed parameter server is configured to update a parameter of an initial model based on the gradient information of the distributed trainer to obtain a trained target model. This embodiment provides an apparatus training a distributed model. In the model training process, the modules are independent of each other and run in cooperation, improving a model training speed.