[{"data":1,"prerenderedAt":391},["ShallowReactive",2],{"content-query-iLcTYFHgfK":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":385,"_id":386,"_source":387,"_file":388,"_stem":389,"_extension":390},"/version-updates/zh/237","zh",false,"","MindSpore新版本v0.6.0-beta解读","大家好，又到了MindSpore每月更新跟大家见面的时间，这次发布新版本的是v0.6.0-beta（后简称v0.6）版本,让我们一起来看看吧！","2020-07-31","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/09/21/98222cafe6f8499db68d62500c244866.png","version-updates",{"type":14,"children":15,"toc":376},"root",[16,24,30,35,43,48,55,74,79,90,96,101,106,111,116,139,146,150,159,164,169,174,181,186,193,198,203,212,217,226,236,244,253,261,270,273,280,298,306,311,316,321,326,331,338,343,354,364,369],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"mindspore新版本v060-beta解读",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":29},"大家好，又到了MindSpore每月更新跟大家见面的时间，这次发布新版本的是v0.6.0-beta（后简称v0.6）版本。",{"type":17,"tag":25,"props":31,"children":32},{},[33],{"type":23,"value":34},"Follow小编的版本解读来品品，经过一个月的开发后，MindSpore又有了什么新的关键特性呢。、",{"type":17,"tag":25,"props":36,"children":37},{},[38],{"type":17,"tag":39,"props":40,"children":42},"img",{"alt":7,"src":41},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/07/31/40cb21582e11450aadb332567b627441.gif",[],{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":23,"value":47},"老规矩先上视频，看视频发现MindSporev0.6版本新特性",{"type":17,"tag":49,"props":50,"children":52},"h2",{"id":51},"mindspore-serving",[53],{"type":23,"value":54},"MindSpore Serving",{"type":17,"tag":56,"props":57,"children":58},"ul",{},[59,65],{"type":17,"tag":60,"props":61,"children":62},"li",{},[63],{"type":23,"value":64},"训练完成后无缝对接推理服务",{"type":17,"tag":60,"props":66,"children":67},{},[68,70],{"type":23,"value":69},"支持推理模型的预加载，获得更佳性能的推理服务",{"type":17,"tag":39,"props":71,"children":73},{"alt":7,"src":72},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/07/31/87192ce664f04acebd301a12bd1a6eba.jpg",[],{"type":17,"tag":25,"props":75,"children":76},{},[77],{"type":23,"value":78},"官网教程：",{"type":17,"tag":25,"props":80,"children":81},{},[82],{"type":17,"tag":83,"props":84,"children":88},"a",{"href":85,"rel":86},"https://www.mindspore.cn/tutorial/zh-CN/r0.6/advanced_use/serving.html",[87],"nofollow",[89],{"type":23,"value":85},{"type":17,"tag":49,"props":91,"children":93},{"id":92},"ps分布式训练",[94],{"type":23,"value":95},"PS分布式训练",{"type":17,"tag":25,"props":97,"children":98},{},[99],{"type":23,"value":100},"分布式训练作为MindSpore最受瞩目的特性之一，MindSpore团队在v0.6版本里又添加了新的内容——PS分布式训练。",{"type":17,"tag":25,"props":102,"children":103},{},[104],{"type":23,"value":105},"Parameter Server(参数服务器)是分布式训练中一种广泛使用的架构，既支持同步SGD，也支持异步SGD的训练算法。在扩展性上，将模型的计算与模型的更新分别部署在Worker和Server两类进程中，使得Worker和Server的资源可以独立地横向扩缩。",{"type":17,"tag":25,"props":107,"children":108},{},[109],{"type":23,"value":110},"另外，在大规模数据中心的环境下，计算设备、网络以及存储经常会出现各种故障而导致部分节点异常，而在参数服务器的架构下，能够较为容易地处理此类的故障而不会对训练中的任务产生影响。",{"type":17,"tag":25,"props":112,"children":113},{},[114],{"type":23,"value":115},"在MindSpore的参数服务器一共包含三个独立的组件，分别是Server、Worker和Scheduler，作用分别是：",{"type":17,"tag":56,"props":117,"children":118},{},[119,124,129,134],{"type":17,"tag":60,"props":120,"children":121},{},[122],{"type":23,"value":123},"Server：保存模型的权重和反向计算的梯度值，并使用优化器通过Worker上传的梯度值对模型进行更新（当前版本仅支持单Server）；",{"type":17,"tag":60,"props":125,"children":126},{},[127],{"type":23,"value":128},"Worker：执行网络的正反向计算，正向计算的梯度值通过Push接口上传至Server中，通过Pull接口把Server更新好的模型下载到Worker本地；",{"type":17,"tag":60,"props":130,"children":131},{},[132],{"type":23,"value":133},"Scheduler：用于建立Server和Worker的通信关系；",{"type":17,"tag":60,"props":135,"children":136},{},[137],{"type":23,"value":138},"相较于同步的AllReduce训练方法，Parameter Server具备着很好的灵活性、可扩展性以及节点容灾的能力。",{"type":17,"tag":25,"props":140,"children":141},{},[142],{"type":17,"tag":39,"props":143,"children":145},{"alt":7,"src":144},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/07/31/d8e5cca2fe464ce6836d793ec12255c5.jpg",[],{"type":17,"tag":25,"props":147,"children":148},{},[149],{"type":23,"value":78},{"type":17,"tag":25,"props":151,"children":152},{},[153],{"type":17,"tag":83,"props":154,"children":157},{"href":155,"rel":156},"https://www.mindspore.cn/tutorial/zh-CN/r0.6/advanced_use/parameter_server_training.html",[87],[158],{"type":23,"value":155},{"type":17,"tag":49,"props":160,"children":162},{"id":161},"超大规模稀疏特征训练",[163],{"type":23,"value":161},{"type":17,"tag":25,"props":165,"children":166},{},[167],{"type":23,"value":168},"像wide&Deep的推荐网络，根据场景不同特征数量可能是千万到百亿的规模，Embedding参数量是GB级到TB级。这么大的参数量，远远超过了设备内存容量，无法直接在设备卡上进行训练。",{"type":17,"tag":25,"props":170,"children":171},{},[172],{"type":23,"value":173},"为此，MindSpore在全下沉模式（所有计算都在Device上）之外，提供了Host-Device混合执行模式，即Embedding部署在host测，DNN和通信下沉到Device侧，来解决embedding容量的问题，发挥昇腾芯片算力；并在此基础上实现了混合并行（Embedding模型并行，DNN数据并行）实现超大规模Embedding的高性能训练。",{"type":17,"tag":25,"props":175,"children":176},{},[177],{"type":17,"tag":39,"props":178,"children":180},{"alt":7,"src":179},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/08/01/055450dde234456fb3ffe820da4f03a4.jpg",[],{"type":17,"tag":25,"props":182,"children":183},{},[184],{"type":23,"value":185},"对于第二种模式，目前可以支持500GB的模型，使用creteo数据集训练的性能和精度测试结果如下表：",{"type":17,"tag":25,"props":187,"children":188},{},[189],{"type":17,"tag":39,"props":190,"children":192},{"alt":7,"src":191},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/08/01/ca2c89db2d6d4c14a58478d7112b26c0.jpg",[],{"type":17,"tag":49,"props":194,"children":196},{"id":195},"总结",[197],{"type":23,"value":195},{"type":17,"tag":25,"props":199,"children":200},{},[201],{"type":23,"value":202},"以上就是MindSpore v0.6版本关键特性。当然新版本还有不少特性，详细的特性，可以参考下方RELEASE.md文档：",{"type":17,"tag":25,"props":204,"children":205},{},[206],{"type":17,"tag":83,"props":207,"children":210},{"href":208,"rel":209},"https://gitee.com/mindspore/mindspore/blob/r0.6/RELEASE.md",[87],[211],{"type":23,"value":208},{"type":17,"tag":25,"props":213,"children":214},{},[215],{"type":23,"value":216},"MindSpore保持持续的更新完善，未来会有更多实用的特性发布，非常欢迎大家下载并体验MindSpore，参与到MindSpore开源社区建设中来。",{"type":17,"tag":25,"props":218,"children":219},{},[220],{"type":17,"tag":221,"props":222,"children":223},"strong",{},[224],{"type":23,"value":225},"官方网站：",{"type":17,"tag":25,"props":227,"children":228},{},[229],{"type":17,"tag":83,"props":230,"children":233},{"href":231,"rel":232},"https://www.mindspore.cn/",[87],[234],{"type":23,"value":235},"https://www.mindspore.cn",{"type":17,"tag":25,"props":237,"children":238},{},[239],{"type":17,"tag":221,"props":240,"children":241},{},[242],{"type":23,"value":243},"Gitee：",{"type":17,"tag":25,"props":245,"children":246},{},[247],{"type":17,"tag":83,"props":248,"children":251},{"href":249,"rel":250},"https://gitee.com/mindspore/mindspore",[87],[252],{"type":23,"value":249},{"type":17,"tag":25,"props":254,"children":255},{},[256],{"type":17,"tag":221,"props":257,"children":258},{},[259],{"type":23,"value":260},"GitHub：",{"type":17,"tag":25,"props":262,"children":263},{},[264],{"type":17,"tag":83,"props":265,"children":268},{"href":266,"rel":267},"https://github.com/mindspore-ai",[87],[269],{"type":23,"value":266},{"type":17,"tag":49,"props":271,"children":272},{"id":7},[],{"type":17,"tag":25,"props":274,"children":275},{},[276],{"type":17,"tag":39,"props":277,"children":279},{"alt":7,"src":278},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/08/01/56e4d7c3ec814ec9bf3335b36618a38b.gif",[],{"type":17,"tag":25,"props":281,"children":282},{},[283],{"type":17,"tag":221,"props":284,"children":285},{},[286,288,296],{"type":23,"value":287},"接下来",{"type":17,"tag":221,"props":289,"children":290},{},[291],{"type":17,"tag":221,"props":292,"children":293},{},[294],{"type":23,"value":295},"隆重介绍",{"type":23,"value":297},"即将和大家见面的",{"type":17,"tag":25,"props":299,"children":300},{},[301],{"type":17,"tag":221,"props":302,"children":303},{},[304],{"type":23,"value":305},"第三期MindSpore 两日集训营",{"type":17,"tag":25,"props":307,"children":308},{},[309],{"type":23,"value":310},"集训时间为：8月8日-8月9日两天",{"type":17,"tag":25,"props":312,"children":313},{},[314],{"type":23,"value":315},"报名从今天（7月31日）开始",{"type":17,"tag":25,"props":317,"children":318},{},[319],{"type":23,"value":320},"长按下方图片添加小助手微信",{"type":17,"tag":25,"props":322,"children":323},{},[324],{"type":23,"value":325},"*备注：集训",{"type":17,"tag":25,"props":327,"children":328},{},[329],{"type":23,"value":330},"即可报名",{"type":17,"tag":25,"props":332,"children":333},{},[334],{"type":17,"tag":39,"props":335,"children":337},{"alt":7,"src":336},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/08/01/2b05438596fb451d91c5b12fb0fe6dce.jpg",[],{"type":17,"tag":25,"props":339,"children":340},{},[341],{"type":23,"value":342},"MindSpore官方资料",{"type":17,"tag":25,"props":344,"children":345},{},[346,348],{"type":23,"value":347},"GitHub:",{"type":17,"tag":83,"props":349,"children":352},{"href":350,"rel":351},"https://github.com/mindspore-ai/mindspore",[87],[353],{"type":23,"value":350},{"type":17,"tag":25,"props":355,"children":356},{},[357,359],{"type":23,"value":358},"Gitee:",{"type":17,"tag":83,"props":360,"children":362},{"href":249,"rel":361},[87],[363],{"type":23,"value":249},{"type":17,"tag":25,"props":365,"children":366},{},[367],{"type":23,"value":368},"官方QQ群: 871543426",{"type":17,"tag":25,"props":370,"children":371},{},[372],{"type":17,"tag":39,"props":373,"children":375},{"alt":7,"src":374},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2020/08/01/55a727bcc2284599afd2ef476fd03cbb.jpg",[],{"title":7,"searchDepth":377,"depth":377,"links":378},4,[379,381,382,383,384],{"id":51,"depth":380,"text":54},2,{"id":92,"depth":380,"text":95},{"id":161,"depth":380,"text":161},{"id":195,"depth":380,"text":195},{"id":7,"depth":380,"text":7},"markdown","content:version-updates:zh:237.md","content","version-updates/zh/237.md","version-updates/zh/237","md",1776506144992]