计算机科学 ›› 2020, Vol. 47 ›› Issue (5): 144-148.doi: 10.11896/jsjkx.190700176
郑哲1,2,3, 胡庆浩2, 刘青山1,3, 冷聪2
ZHENG Zhe1,2,3, HU Qing-hao2, LIU Qing-shan1,3, LENG Cong2
摘要: 近年来,生成对抗网络(Generative Adversarial Networks,GAN)在图像超分辨率、图像生成等许多计算机视觉任务中展现出优异的性能。借助于GPU强大的计算力,人们可以设计计算复杂度更高的GAN网络。然而,对于资源受限的移动端设备,高功耗、计算需求大的GAN将很难被直接部署到实际应用中。得益于神经网络压缩技术取得的巨大进展,将GAN部署到移动端设备成为可能。为此,文中提出一种同时对网络权值和激活进行量化的方案来压缩GAN网络。通过量化敏感性分析发现,与量化分类网络不同,GAN中的量化权重比量化激活更敏感,因此在量化时给予权重更多的量化比特。文中比较了两种评价GAN生成图像的指标即Inception Score(IS)和Fréchet Inception Distance(FID),发现FID更适合评估量化后GAN的性能。基于敏感性分析在Mnist和Celeb-A数据集上进行量化实验,用FID指标来评估量化GAN的性能。实验结果表明:在生成图像质量不下降的情况下,所提方法依然可以取得4倍以上的压缩率,从而有效地解决了GAN的压缩问题。
中图分类号:
[1]GOODFELLOW IAN J.Generative adversarial nets[C]//Proceedings of the Neural Information Processing Systems.2014:2672-2680. [2]KRIZHEVSKY A.ImageNet classification with deep convolutional neural networks[C]//Proceedings of the Neural Information Processing Systems.2012:1106-1114. [3]BROCK A,DONAHUE J,SIMONYAN K.Large Scale GAN Training for High Fidelity Natural Image Synthesis[J].arXiv:1809.11096. [4]WANG P,WANG D,JI Y,et al.QGAN:Quantized Generative Adversarial Networks[J].arXiv:1901.08263. [5]ANGELINE A,CHIANG P,GAIN A,et al.Compressing gans using knowledge distillation[J].arXiv:1902.00159. [6]HEUSEL M,RAMSAUER H,UNTERHINER T,et al.GANs Trained by a TWO Time-Scale Update Rule Converge to a Local Nash Equilibrium[J].arXiv:1706.08500. [7]SALIMANS T,GOODFELLOW I,ZAREMBA W,et al.Im-proved Techniques for Training GANs[C]//Proceedings of the Neural Information Processing Systems.2016:2234-2242. [8]MIRZA M,OSINDERO S.Conditional generative adversarialnets[J].arXiv:1411.1784. [9]KARRAS T,AILA T,LAINE S,et al.Progressivegrowing of gans for improved quality,stability,andvariation[J].arXiv:1710.10196. [10]RADFORD A,METZ L,CHINTALA S.Unsupervised representation learning with deep convolutional generative adversarial networks[J].arXiv:1511.06434. [11]ARJOVSKY M,CHINTALA S,BOTTOU L.Wasserstein gan[J].arXiv:1701.07875. [12]GULRAJANI I,AHMED F,ARJOVSKY M,et al.CImproved training of wasserstein gans[C]//Proceedings of the Neural Information Processing Systems.2017:5767-5777. [13]HUBARA I,COURBARIAUX M,SOUDRY D,et al.Binarized neural networks[C]//Proceedings of Neural Information Processing Systems.2016:4107-4115. [14]COURBARIAUX M,BENGIO Y,DAVID J.Binaryconnect:Training deep neural networks with binary weights duringpro-pagations[C]//Proceedings of the Neural Information Proces-sing.2015:3123-3131. [15]RASTEGARI M.XNOR-Net:ImageNet Classification Using Binary Convolutional Neural Networks[C]//Proceedings of the European Conference on Computer Vision.2016:525-542. [16]HU Q H,WANG P S,CHENG J.From Hashing to CNNs:Training Binary Weight Networks via Hashing[C]//Procee-dings of theAAAI Conferenceon Artificial Intelligence.2018:3247-3254. [17]LIN X,ZHAO C,PAN W.Towards accurate binary convolutional neural network.[C]//Proceedings of the Neural Information Processing Systems.Montreal,2017:345-353. [18]ZHOU A J,YAO A B,GUO Y W,et al.Incremental Network Quantization:Towards Lossless CNNs with Low-Precision Weights[J].arXiv:1702.03044. [19]LI Z,NI B,ZHANG W,et al.Performance guaranteed net-workacceleration via high-order residual quantization[C]//Proceedings of the International Conference on Computer Vision.2017:2584-2592. [20]FROMM J,PATEL S,PHILIPOSE M.Heterogeneous Bitwidth Binarization in Convolutional Neural Networks[C]//Procee-dings of the Neural Information Processing Systems.2018:4006-4015. [21]WANG K,LIU Z J,LIN Y J,et al.HAQ:Hardware-Aware Automated Quantization[J].arXiv:1811.08886. [22]ZHOU S C,WU Y X,NI Z K,et al.DoReFa-Net:Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients[J].arXiv:1606.06160. [23]ZHANG D Q,YANG J L,YE D Q Z,et al.LQ-Nets:Learned Quantization for Highly Accurate and Compact Deep Neural Networks[C]//Proceedings of European Conference on Computer Vision.2018. [24]ZHUANG B H,SHEN C H,TAN M K,et al.Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation[J].arXiv:1811.10413. [25]CAI Z,HE X,SUN J,et al.Deep learning with low precision by halfwavegaussian quantization[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.New York:IEEE Press,2017:5918-5926. [26]BENGIO Y,LÉONARD N,COURVILLE A.Estimating orpropagating gradientsthrough stochastic neurons for conditional computation[J].arXiv:1308.3432. [27]SHANE B,RISHI S.A Note on the Inception Score[J].arXiv:1801.01973. |
[1] | 张佳, 董守斌. 基于评论方面级用户偏好迁移的跨领域推荐算法 Cross-domain Recommendation Based on Review Aspect-level User Preference Transfer 计算机科学, 2022, 49(9): 41-47. https://doi.org/10.11896/jsjkx.220200131 |
[2] | 王冠宇, 钟婷, 冯宇, 周帆. 基于矢量量化编码的协同过滤推荐方法 Collaborative Filtering Recommendation Method Based on Vector Quantization Coding 计算机科学, 2022, 49(9): 48-54. https://doi.org/10.11896/jsjkx.210700109 |
[3] | 孙奇, 吉根林, 张杰. 基于非局部注意力生成对抗网络的视频异常事件检测方法 Non-local Attention Based Generative Adversarial Network for Video Abnormal Event Detection 计算机科学, 2022, 49(8): 172-177. https://doi.org/10.11896/jsjkx.210600061 |
[4] | 戴朝霞, 李锦欣, 张向东, 徐旭, 梅林, 张亮. 基于DNGAN的磁共振图像超分辨率重建算法 Super-resolution Reconstruction of MRI Based on DNGAN 计算机科学, 2022, 49(7): 113-119. https://doi.org/10.11896/jsjkx.210600105 |
[5] | 郝强, 李杰, 张曼, 王路. 基于改进YOLOv3的空间非合作目标部件识别算法 Spatial Non-cooperative Target Components Recognition Algorithm Based on Improved YOLOv3 计算机科学, 2022, 49(6A): 358-362. https://doi.org/10.11896/jsjkx.210700048 |
[6] | 尹文兵, 高戈, 曾邦, 王霄, 陈怡. 基于时频域生成对抗网络的语音增强算法 Speech Enhancement Based on Time-Frequency Domain GAN 计算机科学, 2022, 49(6): 187-192. https://doi.org/10.11896/jsjkx.210500114 |
[7] | 徐辉, 康金梦, 张加万. 基于特征感知的数字壁画复原方法 Digital Mural Inpainting Method Based on Feature Perception 计算机科学, 2022, 49(6): 217-223. https://doi.org/10.11896/jsjkx.210500105 |
[8] | 高志宇, 王天荆, 汪悦, 沈航, 白光伟. 基于生成对抗网络的5G网络流量预测方法 Traffic Prediction Method for 5G Network Based on Generative Adversarial Network 计算机科学, 2022, 49(4): 321-328. https://doi.org/10.11896/jsjkx.210300240 |
[9] | 黎思泉, 万永菁, 蒋翠玲. 基于生成对抗网络去影像的多基频估计算法 Multiple Fundamental Frequency Estimation Algorithm Based on Generative Adversarial Networks for Image Removal 计算机科学, 2022, 49(3): 179-184. https://doi.org/10.11896/jsjkx.201200081 |
[10] | 李建, 郭延明, 于天元, 武与伦, 王翔汉, 老松杨. 基于生成对抗网络的多目标类别对抗样本生成算法 Multi-target Category Adversarial Example Generating Algorithm Based on GAN 计算机科学, 2022, 49(2): 83-91. https://doi.org/10.11896/jsjkx.210800130 |
[11] | 谈馨悦, 何小海, 王正勇, 罗晓东, 卿粼波. 基于Transformer交叉注意力的文本生成图像技术 Text-to-Image Generation Technology Based on Transformer Cross Attention 计算机科学, 2022, 49(2): 107-115. https://doi.org/10.11896/jsjkx.210600085 |
[12] | 陈贵强, 何军. 自然场景下遥感图像超分辨率重建算法研究 Study on Super-resolution Reconstruction Algorithm of Remote Sensing Images in Natural Scene 计算机科学, 2022, 49(2): 116-122. https://doi.org/10.11896/jsjkx.210700095 |
[13] | 石达, 芦天亮, 杜彦辉, 张建岭, 暴雨轩. 基于改进CycleGAN的人脸性别伪造图像生成模型 Generation Model of Gender-forged Face Image Based on Improved CycleGAN 计算机科学, 2022, 49(2): 31-39. https://doi.org/10.11896/jsjkx.210600012 |
[14] | 唐雨潇, 王斌君. 基于深度生成模型的人脸编辑研究进展 Research Progress of Face Editing Based on Deep Generative Model 计算机科学, 2022, 49(2): 51-61. https://doi.org/10.11896/jsjkx.210400108 |
[15] | 张玮琪, 汤轶丰, 李林燕, 胡伏原. 基于场景图的段落生成序列图像方法 Image Stream From Paragraph Method Based on Scene Graph 计算机科学, 2022, 49(1): 233-240. https://doi.org/10.11896/jsjkx.201100207 |
|