Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions
In this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vec...
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10414068/ |
| _version_ | 1797311620656922624 |
|---|---|
| author | Yan-Jen Huang Hsin-Lung Wu Ching-Chen |
| author_facet | Yan-Jen Huang Hsin-Lung Wu Ching-Chen |
| author_sort | Yan-Jen Huang |
| collection | DOAJ |
| description | In this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vectors. Then it eliminates most basis block vectors and their corresponding coefficients to obtain an approximated convolutional kernel. Moreover, the proposed BlkSConv operation can be efficiently realized via a combination of pointwise and group-wise convolutions. Thus the constructed networks have smaller model size and fewer multiply-adds operations while keeping comparable prediction accuracy. We also develop a hyperparameter search framework based on principal component analysis (PCA) to determine a qualified hyperparameter setting of the block depth and number of basis block vectors. By this search framework, we construct networks which achieve nice prediction performance while simultaneously satisfying the constraints of model size and model efficiency. Our code, data, and models are available at <uri>https://github.com/yanjenhuang/blksconv</uri>. |
| first_indexed | 2024-03-08T02:03:20Z |
| format | Article |
| id | doaj.art-00b4ee3eeade4b1f9aff567d5c0ab70c |
| institution | Directory Open Access Journal |
| issn | 2169-3536 |
| language | English |
| last_indexed | 2024-03-08T02:03:20Z |
| publishDate | 2024-01-01 |
| publisher | IEEE |
| record_format | Article |
| series | IEEE Access |
| spelling | doaj.art-00b4ee3eeade4b1f9aff567d5c0ab70c2024-02-14T00:01:46ZengIEEEIEEE Access2169-35362024-01-0112215592156810.1109/ACCESS.2024.335862010414068Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard ConvolutionsYan-Jen Huang0https://orcid.org/0009-0009-4659-8670Hsin-Lung Wu1https://orcid.org/0000-0002-1129-3668 Ching-Chen2Department of Computer Science and Information Engineering, National Taipei University, New Taipei City, TaiwanDepartment of Computer Science and Information Engineering, National Taipei University, New Taipei City, TaiwanDepartment of Computer Science and Information Engineering, National Taipei University, New Taipei City, TaiwanIn this paper, we introduce block-wise separable convolutions (BlkSConv) to replace the standard convolutions for compressing deep CNN models. First, BlkSConv expresses the standard convolutional kernel as an ordered set of block vectors each of which is a linear combination of fixed basis block vectors. Then it eliminates most basis block vectors and their corresponding coefficients to obtain an approximated convolutional kernel. Moreover, the proposed BlkSConv operation can be efficiently realized via a combination of pointwise and group-wise convolutions. Thus the constructed networks have smaller model size and fewer multiply-adds operations while keeping comparable prediction accuracy. We also develop a hyperparameter search framework based on principal component analysis (PCA) to determine a qualified hyperparameter setting of the block depth and number of basis block vectors. By this search framework, we construct networks which achieve nice prediction performance while simultaneously satisfying the constraints of model size and model efficiency. Our code, data, and models are available at <uri>https://github.com/yanjenhuang/blksconv</uri>.https://ieeexplore.ieee.org/document/10414068/Convolutional neural networkblock-wise separable convolutionnetwork architecture search |
| spellingShingle | Yan-Jen Huang Hsin-Lung Wu Ching-Chen Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions IEEE Access Convolutional neural network block-wise separable convolution network architecture search |
| title | Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions |
| title_full | Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions |
| title_fullStr | Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions |
| title_full_unstemmed | Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions |
| title_short | Block-Wise Separable Convolutions: An Alternative Way to Factorize Standard Convolutions |
| title_sort | block wise separable convolutions an alternative way to factorize standard convolutions |
| topic | Convolutional neural network block-wise separable convolution network architecture search |
| url | https://ieeexplore.ieee.org/document/10414068/ |
| work_keys_str_mv | AT yanjenhuang blockwiseseparableconvolutionsanalternativewaytofactorizestandardconvolutions AT hsinlungwu blockwiseseparableconvolutionsanalternativewaytofactorizestandardconvolutions AT chingchen blockwiseseparableconvolutionsanalternativewaytofactorizestandardconvolutions |