深度学习在线下场景的探索与实践-云栖大会（38页）.pdf

1、云端3IC4min W+.=G=0 min W.C=1,15,=W+2 2+=+min 22.,=W+2 2+1:=argmin,+1:=argmin+1,+1:=+1+1min2=+1+,.6Continuous solutionDiscrete solution7Binary：C=1,1Ternary：C=1,0,1Int3：C=2,1,0,1,2,C=4,2,1,0,1,2,4,C=8,4,2,1,0,1,2,4,FeatureMap量化：minii 2.int4量化AlexNet9MethodMethod#Bits#BitsTopTop-1 1 Acc.Acc.TopTop-5 5 A

2、cc.Acc.Binary(ours)10.6870.886BWN10.6390.851Ternary(ours)20.7250.907TWN20.6560.8653bit（ours）30.7490.921Full-precision320.7530.9223 3 BitsBitsFullFullmAP0.7760.77810Prune 60%Weights(a)(b)(c)70%90%.Starting PointAlexnetGoogleNetDC：Deep Compression:Compressing Deep Neural Networks with Pruning,Trained

3、Quantization and Huffman CodingDNS：Dynamic Network Surgery for Efficient DNNs精度无损的情况下：+2+1-2-113xyd(x,y)p=1/d x,y2x,yO(log(n)。141/d21516IPIRC（ICCV）2019 第一名Low Power Image Recognition Challenge1718Comparison with the state-of-the-art methods on COCO test-dev set.19高斯压缩工具TS结构实现无感知压缩202122参数稀疏：90%+量化4,2,1,0分支23170倍加速比0.173ms0.173msGPU29.9ms242527人机对比（精度）“天巡”智能算法人工标记目标识别（*市大棚）9 90%0%60%*市疑似违章识别78%78%34%路网提取排行榜第一地物分类排行榜第一遥感目标检测排行榜第一人机对比结果28前一年卫片后一年卫片及变化前一年卫片后一年卫片新增建筑物识别路网提取结果地物分类北京市大棚识别局部放大图遥感密集目标检测前一年卫片后一年卫片新增建筑物识别结果类别A-背包、行李箱内危险品检测34