閱讀全文 | |
篇名 |
Variational Automatic Channel Pruning Algorithm Based on Structure Optimization for Convolutional Neural Networks
|
---|---|
並列篇名 | Variational Automatic Channel Pruning Algorithm Based on Structure Optimization for Convolutional Neural Networks |
作者 | Shuo Han、Yufei Zhan、Xingang Liu |
英文摘要 | Channel pruning has achieved remarkable success in solving the significant computation and memory consumption in Convolutional Neural Networks (CNN). Most existing methods measure the importance of channels with manually designed algorithms and pruning unimportant channels rely on heuristics or expertise during the processing, which are labourious and subjective. In this paper, we proposed a Variational Automatic Channel Pruning Algorithm based on structure optimization (VA-CPSO) which can automatically optimize channel numbers via channel scales in end-toend manner through variational inference. Firstly, a weights generator controlled by channel scales is built to produce weights for various pruned structure of CNN. And then, the channel scales with truncated factorized log-uniform prior and log-normal posterior are optimized by variational inference for optimal pruning structure. Meanwhile, parameters of the weights generator are optimized synchronously. Finally, the acquired optimal structure and corresponding generated weights are deployed in the pruned CNN for further training to achieve high-performance compression. The experimental results demonstrate that our proposed VACPSO acquires better compression performance compared to existing pruning algorithms. The VA-CPSO achieve a compression of 34.60×, 4.28×, 1.96× and a speedup of 28.20×, 2.03×, 2.02× on LeNet-5, VGGNet, and ResNet-110 with no more than 0.5% loss of accuracy. |
起訖頁 | 339-351 |
關鍵詞 | Automatic channel pruning、Variational inference、Structure optimization、CNN compression、Truncated distributions |
刊名 | 網際網路技術學刊 |
期數 | 202103 (22:2期) |
出版單位 | 台灣學術網路管理委員會 |
DOI |
|
QR Code | |
該期刊 上一篇
| Blockchain-enabled Charging Scheduling for Unmanned Vehicles in Smart Cities |
該期刊 下一篇
| Energy Conserving Forepart Detection Scheme with Dynamic Compressive Measurements Based on Compressive Sensing for WVSN |