A model pruning method, an image processing method and related devices
CN116128044BActive Publication Date: 2026-08-28JILUO TECH (SHANGHAI) CO LTD
Patent Information
- Application Number
- CN202211600306.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-13
- Publication Date
- 2026-08-28
- Estimated Expiration
- 2042-12-13
AI Technical Summary
Technical Problem
当前模型剪枝分为迭代式剪枝和一次剪枝(one-shot剪枝)，两者在分析剪枝配置时通常采用贪心法确定剪枝配置，该方法无法达到全局最优，且对复杂的神经网络模型来说剪枝效果不够好
Benefits of technology
[0048]本发明提供的模型剪枝方法、图像处理方法及相关装置，其中，模型剪枝方法通过先对待剪枝模型进行预剪枝获得目标剪枝配置，再根据目标剪枝配置对待剪枝模型进行一次剪枝，从而获得目标模型。其中，预剪枝过程是先根据剪枝目标对待剪枝模型进行均匀剪枝获得初始剪枝配置，再进行迭代剪枝获得目标剪枝配置，从全局最优的角度设计预剪枝过程，因而获得全局最优的剪枝配置，使得复杂的待剪枝模型具有较好的剪枝效果。
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN116128044B_ABST
Abstract
The application provides a model pruning method, an image processing method and related devices, including: obtaining a to-be-pruned model and a corresponding pruning target, wherein the to-be-pruned model is obtained based on image data training; performing pre-pruning on the to-be-pruned model according to the pruning target to obtain a target pruning configuration; performing one-time pruning on the to-be-pruned model according to the target pruning configuration and the pruning target to obtain a target model; wherein the pre-pruning of the to-be-pruned model according to the pruning target to obtain the target pruning configuration includes: uniformly pruning the channels of each network layer in the to-be-pruned model according to the pruning target to obtain an initial pruning configuration; based on the pruning target and the initial pruning configuration, iteratively pruning the channels of all network layers in the to-be-pruned model to obtain the target pruning configuration. The application starts from the global optimum and has better pruning effect.
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Method, device and equipment for identifying object category in image and storage medium
CN114612719A