CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习

4星(超过85%的资源)
所需积分/C币:9 2016-05-16 18:37:02 1.21MB PDF
38
收藏 收藏
举报

该文档来自CCTC 2016中国云计算技术大会。聚效广告技术经理 & DMLC成员、MXNet committer刘忆智发表的题为“Beyond MLLibScale up Advanced Machine Learning on Spark”的主题演讲,欢迎下载!
Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 Computational Mode task task task mode direction direction direction pdate data barrier aggregated direction Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 Advanced ml algorithms mIc XGBoost XGBoost is an optimized distributed gradient boosting library designed to be highly efficient, flexible and portable dmlc MXNet is a deep learning framework designed for both efficiency and flexibility monet Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 Deep Learning 门 02 p(cat) 85 p(dog) Q身。是是 督普 1 deep learning"trend in the past 10 years 2013 215 Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 Deep Learning 4恩 19c 1 hot\/ log \dense 12 13 13 P24 dense dens 27 100 192 128 Max MER pooling 24 204B stride Max pooling Doolin Hard to define the network Huge computational cost Convolution layers Fully connected layers Memory limit Way to distribute training process Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 MXNet overview c/C++ Python‖R…Juia ND Symbolic EXpr KV Array Binder Store BLAS Dep Engine Comm CPU‖GPU‖ Android - iOS Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 Train Deep network on MXNet (declaratively) define layers model definition val data Symbol Variable( data") val fcl Symbol FullyConnected(name =fcl )(Map( data" -> data,num hidden"-> 128)) val actl Symbol Activation(name ="relul )(Map( data-> fcl,"act type relu)) val act2= symbol Activation(name ="relu2")(Map( data"-> fc2, act type er val fc2 Symbol FullyConnected(name =fc2 )(Map( data"-> actl,"num hidde n"->64)) >"re1u")) val fc3 Symbol FullyConnected(name =fc3 )(Map( data"-> act2,num hidden"-> 10) val mlp Symbol Softmaxoutput(name =sm")(Map( data fc3)) setup model and fit the training set Ices here, e.g. Context. gpu(0, 1, 2, 3 val model FeedForward. newBuilder(mlp) setContext(Context. cpu() setNumEpoch(10) setoptimizer(new SGD (learningRate =0.lf, momentum =0.9f, wd =0.0001f) setrrainData(trainDatafter) setEvalData(valDataIter) build val probArrays model predict(valDataIter) in this case, we do not have multiple outputs require (probArrays length = 1 val prob probArrays(o) user defined optimizer get predicted labels val py NDArray. argmaxChannel(prob) // deal with predicted labels py Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 MXNet benchmarks Compare mXNet to others Internal memory usage of MXNet on a single forward-backward under various allocation strategies performance 1000 TensorFlow na e TaME Calle inplace inplace Torch? 86420 co-share co-share MANet inplace co- share m inplace&oo-share 60D 400 EE 200 420 alexnet googlenet vgg alexnet googlenet vgg alexnet googlenet veg Beyond MLLib: Scale up Advanced Machine Learning on Spark CCTc2016中计技水大盒 GPU Support NDArray is imperative specify device here val weight =NDArray empty( Shape(3, 2), Context. gpu(o)) weight - eta *(grad +lambda weight), val weighton Cpu weight copy To( Context. cpu) Copy to another device One code for both cpu and gpu (shadow) translates at compile time template<typename xpu> void UpdateSGD(Tensor<xpu, 2> weight, const Tensor< xpu, 2>&grad float eta, float lambda)i welght eta *(grad +lambda weight);

...展开详情
试读 36P CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习
立即下载 身份认证后 购VIP低至7折
一个资源只可评论一次,评论内容不能少于5个字
paxinla 感谢楼主的分享!
2016-10-23
回复
您会向同学/朋友/同事推荐我们的CSDN下载吗?
谢谢参与!您的真实评价是我们改进的动力~
  • 分享王者

关注 私信
上传资源赚钱or赚积分
最新推荐
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习 9积分/C币 立即下载
1/36
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第1页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第2页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第3页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第4页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第5页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第6页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第7页
CCTC 2016 聚效广告刘忆智:超越MLLib: 通过XGBoost/MXNet 看Spark上的前沿(深度)机器学习第8页

试读结束, 可继续读3页

9积分/C币 立即下载