PENNI: Pruned Kernel Sharing for Efficient CNN Inference

Shiyu Li, Edward Hanson, Hai Li, Yiran Chen
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:5863-5873, 2020.

Abstract

Although state-of-the-art (SOTA) CNNs achieve outstanding performance on various tasks, their high computation demand and massive number of parameters make it difficult to deploy these SOTA CNNs onto resource-constrained devices. Previous works on CNN acceleration utilize low-rank approximation of the original convolution layers to reduce computation cost. However, these methods are very difficult to conduct upon sparse models, which limits execution speedup since redundancies within the CNN model are not fully exploited. We argue that kernel granularity decomposition can be conducted with low-rank assumption while exploiting the redundancy within the remaining compact coefficients. Based on this observation, we propose PENNI, a CNN model compression framework that is able to achieve model compactness and hardware efficiency simultaneously by (1) implementing kernel sharing in convolution layers via a small number of basis kernels and (2) alternately adjusting bases and coefficients with sparse constraints. Experiments show that we can prune 97% parameters and 92% FLOPs on ResNet18 CIFAR10 with no accuracy loss, and achieve a 44% reduction in run-time memory consumption and a 53% reduction in inference latency.

Cite this Paper


BibTeX
@InProceedings{pmlr-v119-li20d, title = {{PENNI}: Pruned Kernel Sharing for Efficient {CNN} Inference}, author = {Li, Shiyu and Hanson, Edward and Li, Hai and Chen, Yiran}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {5863--5873}, year = {2020}, editor = {Hal Daumé III and Aarti Singh}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/li20d/li20d.pdf}, url = { http://proceedings.mlr.press/v119/li20d.html }, abstract = {Although state-of-the-art (SOTA) CNNs achieve outstanding performance on various tasks, their high computation demand and massive number of parameters make it difficult to deploy these SOTA CNNs onto resource-constrained devices. Previous works on CNN acceleration utilize low-rank approximation of the original convolution layers to reduce computation cost. However, these methods are very difficult to conduct upon sparse models, which limits execution speedup since redundancies within the CNN model are not fully exploited. We argue that kernel granularity decomposition can be conducted with low-rank assumption while exploiting the redundancy within the remaining compact coefficients. Based on this observation, we propose PENNI, a CNN model compression framework that is able to achieve model compactness and hardware efficiency simultaneously by (1) implementing kernel sharing in convolution layers via a small number of basis kernels and (2) alternately adjusting bases and coefficients with sparse constraints. Experiments show that we can prune 97% parameters and 92% FLOPs on ResNet18 CIFAR10 with no accuracy loss, and achieve a 44% reduction in run-time memory consumption and a 53% reduction in inference latency.} }
Endnote
%0 Conference Paper %T PENNI: Pruned Kernel Sharing for Efficient CNN Inference %A Shiyu Li %A Edward Hanson %A Hai Li %A Yiran Chen %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-li20d %I PMLR %P 5863--5873 %U http://proceedings.mlr.press/v119/li20d.html %V 119 %X Although state-of-the-art (SOTA) CNNs achieve outstanding performance on various tasks, their high computation demand and massive number of parameters make it difficult to deploy these SOTA CNNs onto resource-constrained devices. Previous works on CNN acceleration utilize low-rank approximation of the original convolution layers to reduce computation cost. However, these methods are very difficult to conduct upon sparse models, which limits execution speedup since redundancies within the CNN model are not fully exploited. We argue that kernel granularity decomposition can be conducted with low-rank assumption while exploiting the redundancy within the remaining compact coefficients. Based on this observation, we propose PENNI, a CNN model compression framework that is able to achieve model compactness and hardware efficiency simultaneously by (1) implementing kernel sharing in convolution layers via a small number of basis kernels and (2) alternately adjusting bases and coefficients with sparse constraints. Experiments show that we can prune 97% parameters and 92% FLOPs on ResNet18 CIFAR10 with no accuracy loss, and achieve a 44% reduction in run-time memory consumption and a 53% reduction in inference latency.
APA
Li, S., Hanson, E., Li, H. & Chen, Y.. (2020). PENNI: Pruned Kernel Sharing for Efficient CNN Inference. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:5863-5873 Available from http://proceedings.mlr.press/v119/li20d.html .

Related Material