legonet efficient convolutional neural networks with lego
play

LegoNet: Efficient Convolutional Neural Networks with Lego Filters - PowerPoint PPT Presentation

LegoNet: Efficient Convolutional Neural Networks with Lego Filters Zhaohui Yang 1,2,* Yunhe Wang 2 Han3ng Chen 1,2,* Chuanjian Liu 2 Boxin Shi 3,4 Chao Xu 1 Chunjing Xu 2 Chang Xu 5 1 Laboratory of Machine Percep3on (Ministry of Educa3on), Peking


  1. LegoNet: Efficient Convolutional Neural Networks with Lego Filters Zhaohui Yang 1,2,* Yunhe Wang 2 Han3ng Chen 1,2,* Chuanjian Liu 2 Boxin Shi 3,4 Chao Xu 1 Chunjing Xu 2 Chang Xu 5 1 Laboratory of Machine Percep3on (Ministry of Educa3on), Peking University 2 Huawei Noah’s Ark Lab 3 Peng Cheng Laboratory 4 Na3onal Engineering Laboratory For Video Technology, Peking University 5 School of Computer Science, University of Sydney * This work was down when Zhaohui Yang and Han3ng Chen were interns at Huawei Noah’s Ark Lab

  2. Goal • Mo&va&on Reuse paXerns • Targeted Build efficient CNN using a set of Lego Filters • Lego Filters Standard convolu3on filters are established by a set of shared filters • Op&miza&on End-to-end op3miza3on, Straight Through Es3mator • Efficient Inference Split-Transform-Merge strategy ICML 2019

  3. Lego Filters • Lego Filters B B = {B 1 , …, B m } • Standard convolu&on filters F F = G(B 1 , …, B m ), 4-D tensor G is a genera3on func3on. • Compression condi&on |G| + |B| ≤ |F| • G in LegoNet Combina3on m: the number of Lego Filters B : lego Filters F : standard convolu3on filters G: genera3on func3on ICML 2019

  4. Op&miza&on • Targeted • Op&mize Binary matrix M Float type proxy weight N Straight Through Es3mator (STE) • Op&mize Lego Filters B Standard BP algorithm m: the number of Lego Filters B : Lego Filters M : binary index matrix N : proxy matrix of M ICML 2019

  5. Lego Unit & Efficient Inference • Split • Split input feature maps X • Transform • Convolve feature fragments X = {X 1 , …, X o } with Lego Filters B = {B 1 , …, B m } • Merge • Combine Lego Feature Maps according to learnt combina3on matrix M m: the number of Lego Filters o : split number B : Lego Filters M : binary index matrix ICML 2019

  6. Analysis • Compression • Accelera&on • Condi3on • m ≤ n m: the number of Lego Filters n : the number of output channels o : split number M : binary index matrix ICML 2019

  7. Experiments CIFAR-10 ImageNet Combina3on with coefficients is important while stacking Lego Filters. Given same model size, larger split number o results in higher performance (larger FLOPs) ICML 2019

  8. Conclusion & Future Research • Conclusion • Proposed Lego Filters for construc3ng efficient CNN. • End-to-end op3miza3on. • Split-transform-merge three-stage strategy. • Future Research • Parameter in Parameter (use a set of Lego Filters and a small NN to generate 4-D convolu3on filters) • Global LegoNet (view network parameters as a whole 4-D tensor) ICML 2019

  9. Thanks! ICML 2019

Recommend


More recommend