Reparametric Neural Network Architecture Search for Edge Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network architecture search methods face challenges in achieving high accuracy and real-time detection speed, particularly in mobile devices, due to limitations in branch structures and video memory usage, which hinder the efficient improvement of model performance.
Innovation Solution
An edge calculation-oriented reparametric neural network architecture search method that designs multi-branch blocks, constructs a hypernetwork, and trains it under video memory restrictions to activate and fuse branches, ultimately transforming the network into a single-branch structure for improved reasoning speed and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-branch structure is used to improve feature extraction ability and network accuracy, then network accuracy is improved, but reasoning speed decreases and video memory space increases
Solution Approach 1:
The network is divided into multiple branches during training to improve feature extraction, but each branch processes features independently. During reasoning, these segmented branches are merged into a single unified computation path, eliminating redundant operations while preserving the benefits of multi-branch feature extraction.
Solution Approach 2:
The network structure dynamically changes between training and reasoning phases. During training, the multi-branch structure is active to enhance feature extraction capabilities. During reasoning, the structure transforms into a single-branch equivalent that maintains accuracy while improving speed and reducing memory usage.
2Measurement precision
If multi-branch structure is used to improve network accuracy, then network accuracy is improved, but video memory space required for training increases
Solution Approach 1:
Multiple branch computations are merged into a single unified computation during reasoning. By combining the computational paths of multiple branches into one equivalent operation, the patent reduces video memory requirements while preserving the feature extraction capabilities of the multi-branch structure.
3Speed
If single-branch structure is used to improve reasoning speed, then reasoning speed is improved, but network accuracy decreases
Solution Approach 1:
The patent changes the structural parameters of the network between training and reasoning phases. During training, multi-branch parameters enable high accuracy. During reasoning, these parameters are transformed into a single-branch equivalent that maintains the same accuracy level while improving speed through more efficient computation.
4Quantity of substance
If nonparametric operations such as skip connections are used to reduce parameters, then parameter count is reduced, but reasoning speed is still slowed down
Solution Approach 1:
The patent extracts and eliminates redundant operations from the network structure. By identifying and removing unnecessary computational steps including certain nonparametric operations, the patent achieves both parameter reduction and speed improvement, unlike conventional approaches where skip connections alone are insufficient for speed optimization.
Data Source
AI summary
The present invention discloses an edge calculation-oriented reparametric neural network architecture search method, including the following steps: S1: designing linear operators and multi-branch block structures; S2: constructing a hypernetwork by stacking the multi-branch block structures; S3: training the hypernetwork through a gradient-based first-stage search algorithm; S4: deleting redundant branches in the hypernetwork to construct an optimal subnetwork; S5: converting the multi-branch optimal subnetwork into a single-branch network; and S6: completing task reasoning by using the single-branch network. The method is used to search the neural network structure capable of performing reparameterization, and ensures the reasoning real-time performance and the high efficiency of model operation while ensuring the reasoning precision.