计算机科学 ›› 2025, Vol. 52 ›› Issue (11A): 241100059-7.doi: 10.11896/jsjkx.241100059
张晓玄, 唐小勇
ZHANG Xiaoxuan, TANG Xiaoyong
摘要： 通过鸟瞰全景视角,BEVFormer在自动驾驶领域展现出卓越的性能。然而,在资源受限的设备上,其高内存占用和计算复杂度为实时部署带来了严峻的挑战。BEVFormer中ReLU激活值的分布从零到正无穷,呈现出不均匀的特点,传统量化指标如余弦相似度和均方误差(MSE)无法充分描述这种特性。针对此问题,提出了一种训练后平衡量化策略,该策略专门针对BEVFormer中的线性层和ReLU激活值的量化进行了优化。在线性层权重和输出的量化中采用预定义量化区间,同时对ReLU激活值使用特定区间量化方法,以确保关键值的精确表示。此外,该方法基于Hessian矩阵优化技术实现缩放因子动态调整,利用Hessian矩阵最小化量化误差,并稳定训练过程。实验结果显示,平衡量化策略显著提升了计算效率,同时保证了精度。在nuScenes测试集中,8位量化仅导致NDS下降不到1个百分点,保持了BEVFormer的性能表现。
中图分类号:
| [1]HUANG D Q,HUANG H F,HUANG D Y,et al.A Survey on BEV Perception Learning in Autonomous Driving [J/OL].Computer Engineering and Applications,1-23 [2024-11-06].http://kns.cnki.net/kcms/detail/11.2127.TP.20241031.1529.016.html. [2]MA Y X,WANG T,BAI X Y,et al.Vision-Centric BEV Perception:A Survey[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2024,46(12):10978-10997. [3]LI Z,WANG W,LI H,et al.Bevformer:Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers[C]//European Conference on Computer Vision.Cham:Springer,2022:1-18. [4]YANG C,CHEN Y,TIAN H,et al.Bevformer v2:Adaptingmodern image back-bones to bird’s-eye-view recognition via perspective supervi-sion[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).2023:17830-17839. [5]MA Y,WANG T,BAI X,et al.Vision-centric bev perception:A survey[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2024,46(12):10978-10996. [6]LIU Z,TANG H,AMINI A,et al.Bevfusion:Multi-task multi-sensor fusion with unified bird’s-eye view representation[C]//2023 IEEE Inter-national Conference on Robotics and Automation(ICRA).IEEE,2023:2774-2781. [7]WANG C,WANG Z,XU X,et al.Towards Accurate Post-training Quantization for Diffusion Models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).2024. [8]DIAO H,LI G,XU S,et al.Attention Round for Post-Training Quantization[J].Neurocomputing,2022,521:364-373. [9]BANNER R,NAHSHAN Y,SOUDRY D.Post training 4-bit quantization of convolutional networks for rap-id-deployment[C]//Advances in Neural Information Processing Systems.2019:1-12. [10]CHOUKROUN Y,KRAVCHIK E,YANG F,et al.Low-bitquantization of neural networks for efficient inference[C]//2019 IEEE/CVF International Conference on Computer Vision Workshop(ICCVW).2019:1-12. [11]CHENG Y,WANG D,ZHOU P,et al.A Survey of Model Compression and Acceleration for Deep Neural Networks[J].arXiv:1710.09282,2017. [12]YAO Z,AMINABADI R Y,ZHANG M,et al.ZeroQuant:Efficient and Affordable Post-Training Quantization for Large-Scale Transformers[C]//Advances in Neural Information Processing Systems 35(NeurIPS 2022).2022:27168-27183. [13]VASWANI A,SHAZEER N,PARMAR N,et al.Attention isall you need[C]//Advances in Neural Information Processing Systems.2017. [14]LIU Z,WANG Y,HAN K,et al.Post-training quantization for vision transformer[C]//Advances in Neural Information Processing Systems.2021:28092-28103. [15]YUAN Z,XUE C,CHEN Y,et al.PTQ4ViT:Post-trainingQuantization for Vision Transformers with Twin Uniform Quantization[C]//2021 European Conference on Computer Vision(ECCV).2021:1-12. [16]JIAO J,JIE Z,CHEN S,et al.MSMDFusion:Multi-Depth Seed Based Fusion of LiDAR and Camera Data for Enhanced 3D Object Detection[J].IEEE Transactions on Intelligent Transportation Systems,2023,25:1-12. [17]YIN Z,ZHOU X,KRÄHENBÜHL P.Multimodal Virtual Point 3D Detection:Establishing Fine-Grained Feature Alignment Across Different Modalities[C]//Proceedings of the IEEE International Conference on Computer Vision(ICCV).2021:1-12. [18]BAI H,HU Z,HUANG Q,et al.TransFusion:A Robust Li-DAR-Camera Fusion Method for 3D Object Detection Using Transformers[J].IEEE Transactions on Intelligent Transportation Systems,2022,25:1-12. [19]CHEN L.Dense Projection Fusion(DPFusion):A Multi-Stage,Fine-Grained Fusion Technique for 3D Object Detection[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2024,46:35-56. [20]DOE J.Piecewise Linear Quantization(PWLQ)[J].Neural Networks,2024,58(C):123-145. [21]NAGEL J,LEUNG T,SCHMIDT T,Casale P.AdaRound:Adaptive rounding for post-training quantization[C]//Procee-dings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).2020:1-12. [22]LI B,CHEN Q,XU H,et al.BRECQ:Bit-reconstruction quantization for ultra low-bit quantization aware training[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).2021:1-12. [23]HUBARA I,COURBARIAUX M,SOUDRY D,et al.Block-Wise Mixed-Precision Quantization:Enabling High Efficiency for Practical ReRAM-based DNN Accelerators[J].IEEE Transactions on Circuits and Systems I:Regular Papers,2021,68(11):1-12. [24]BONDARENKO Y,NAGEL M,BLANKEVOORT T.Under-standing and Overcoming the Challenges of Efficient Transformer Quantization[C]//Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.2021:7947-7969. [25]FRANTAR M,HUBARA I,SOUDRY D,et al.CompressingDeep Neural Networks by Combining Weight Pruning with Post-Training Quantization[C]//Proceedings of the 36th Inernational Conference on Neural Information Processing Systems.2022:12760-12773. [26]XIAO G,LIN J,SEZNEC M,et al.SmoothQuant:Accurate and Efficient Post-Training Quantization for Large Language Models[C]//40th International Conference on Machine Learning(ICML).2023:10-12. |
|
||
首页