Deep Learning Model Conversion to Data Flow Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models trained on conventional instruction set architectures face significant challenges when deployed on data flow architectures due to differences in operator granularity and calculation order, hindering efficient computation and application development.
Innovation Solution
A method and apparatus for converting a deep learning model from an instruction set architecture to a data flow architecture by parsing the model into an intermediate representation, converting it, and adjusting it to a customized architecture, allowing the model to operate efficiently on a data flow framework.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a deep learning model is trained on an instruction set architecture, then the model can be developed with high flexibility and ease of operation, but the model cannot be directly deployed on a data flow architecture due to significant differences in operator granularity and data representation
Solution Approach 1:
The patent introduces an intermediate representation as a mediator between the instruction set architecture computation graph and the data flow architecture computation graph. This intermediate representation serves as a bridge that enables conversion between the two different architectural paradigms, allowing models trained on one architecture to be deployed on another without direct compatibility issues.
Solution Approach 2:
The conversion process is segmented into distinct stages: parsing the target deep learning model into an intermediate representation of an instruction set computation graph, converting this intermediate representation into an intermediate representation of a data flow computation graph, adjusting the data flow intermediate representation to a customized architecture, and finally generating the target data flow network model. This segmentation allows each conversion stage to be handled independently and systematically.
2Productivity
If the operator granularity of data flow architecture is increased, then computing efficiency is improved, but the flexibility and ease of operation are reduced compared to instruction set architecture
Solution Approach 1:
The patent creates a computational graph representation that copies the semantic meaning and mathematical operations from the source architecture to the target architecture. By preserving the computational semantics during conversion, the patent maintains the ease of operation from the instruction set architecture while achieving the efficiency benefits of data flow architecture with higher operator granularity.
3Power
If conventional instruction set architecture is used, then development flexibility is maintained, but computing power requirements cannot be met by deep learning applications
Solution Approach 1:
The patent changes key parameters of the computational architecture, specifically transitioning from instruction set architecture with smaller operator granularity to data flow architecture with larger operator granularity. This parameter change enables the system to meet increased computing power requirements while improving execution efficiency through better data representation and computation optimization.
Data Source
AI summary
Provided are a conversion method and apparatus for a deep learning model, a server, and a storage medium. The method includes: parsing a target deep learning model into an intermediate representation of an instruction set computation graph; converting the intermediate representation of the instruction set computation graph into an intermediate representation of a data flow computation graph; adjusting the intermediate representation of the data flow computation graph to an intermediate representation of a customized architecture; and obtaining a converted target data flow network model corresponding to the target deep learning model according to the intermediate representation of the customized architecture.