Neural Network Convolution Kernel Data Manipulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network technologies face inefficiencies in performing convolution operations, particularly when dealing with a large number of channels, which can lead to increased computational requirements and resource utilization.
Innovation Solution
The method involves manipulating kernel data by generating a default tensor and replacing its elements with actual kernel data, optimizing the convolution operation based on the number of input and output channels, and performing the operation in a direction corresponding to the output channels, allowing for efficient group convolution even on general hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional convolution operations are performed on general hardware with a large number of channels, then computational accuracy is maintained, but computational overhead and resource utilization increase significantly
Solution Approach 1:
The patent segments the convolution operation into distinct phases: kernel data manipulation phase and convolution computation phase. By separating these operations and optimizing each independently, the system achieves better resource utilization and reduced computational overhead on general hardware while maintaining accuracy for large channel configurations
Solution Approach 2:
The patent performs preliminary manipulation of kernel data before the actual convolution operation. This includes pre-processing and organizing kernel data in an optimized format that reduces the computational burden during the convolution phase, thereby reducing overall computational overhead while maintaining operational efficiency
2Use of energy by moving object
If group convolution operations are implemented, then computational overhead is reduced, but device complexity increases due to specialized hardware requirements
Solution Approach 1:
The patent implements a universal kernel data manipulation approach that can handle both standard convolution and group convolution operations on general-purpose hardware. By creating a flexible manipulation framework that adapts to different convolution types, the system achieves group convolution efficiency without requiring specialized hardware, thus reducing device complexity while maintaining low computational overhead
3Adaptability or versatility
If the number of input and output channels is increased to improve model capacity, then neural network performance improves, but computational requirements and resource utilization increase
Solution Approach 1:
The patent changes the parameters of kernel data organization and manipulation based on the number of input and output channels. By dynamically adjusting the kernel manipulation strategy according to channel configuration, the system can accommodate high-channel models that improve adaptability and versatility while optimizing resource utilization to prevent excessive computational requirements
Data Source
AI summary
A processor-implemented method of performing a convolution operation is provided. The method includes obtaining input feature map data and kernel data, determine the kernel data based on a number of input channels of the input feature map, a number of output channels of an output feature map, and a number of groups of the input feature map data and a number of groups of the kernel data related to the convolution operation, and performing the convolution operation based on the input feature map data and the determined kernel data.