Computational Graph Segmentation for Dynamic-to-Static Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inefficiency in remote inference due to the inability to directly convert dynamic computational graphs into static computational graphs, necessitating manual modifications.
Innovation Solution
A method and system that divides dynamic computational graphs into subgraphs, converting them into static computational graphs using a deep learning framework interface, and sending the converted subgraphs to a remote device for inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If dynamic computational graph is directly converted to static computational graph, then inference efficiency is improved, but conversion fails in many scenarios requiring manual modification
Solution Approach 1:
The patent segments the dynamic computational graph into multiple static subgraphs by identifying operation nodes and splitting the graph at these points. This segmentation allows each subgraph to be independently converted to a static format, ensuring successful conversion while maintaining the overall computational logic. The segmentation principle resolves the contradiction by breaking down the problematic full-graph conversion into smaller, successfully convertible units.
2Ease of operation
If dynamic computational graph is converted to static computational graph, then remote inference can be performed, but manual modification is required increasing time consumption
Solution Approach 1:
The patent implements a self-service conversion mechanism where the system automatically identifies operation nodes, splits the dynamic computational graph into subgraphs, and converts them to static format without requiring manual intervention. The conversion apparatus autonomously handles the entire process from graph analysis to subgraph generation and conversion, eliminating the need for manual modification while maintaining high conversion success rates.
3Reliability
If dynamic computational graph is divided into subgraphs, then conversion success rate is improved, but system complexity increases
Solution Approach 1:
The patent applies preliminary action by first analyzing the dynamic computational graph to identify operation nodes and determine optimal split points before performing the actual conversion. This preliminary analysis phase prepares the graph structure in advance, making the subsequent conversion process straightforward and systematic. The pre-identification of split points and subgraph boundaries simplifies the overall complexity despite the segmentation process.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
This application discloses a computational graph inference method and system, and a storage medium. The method includes: monitoring a to-be-inferred original-type computational graph; in response to a failure in converting the monitored original-type computational graph into a corresponding target-type computational graph, dividing the original-type computational graph into first-level original-type computational subgraphs; converting the first-level original-type computational subgraph into at least one corresponding target-type computational graph; and generating a to-be-inferred file by using the target-type computational graph corresponding to the first-level original-type computational subgraph, and sending the to-be-inferred file to a remote device. A technical problem of low inference efficiency of computational graphs is resolved in this application.