[{"data":1,"prerenderedAt":181},["ShallowReactive",2],{"content-query-6WR0JpLsB5":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":175,"_id":176,"_source":177,"_file":178,"_stem":179,"_extension":180},"/technology-blogs/zh/1562","zh",false,"","谷歌抛弃 TensorFlow，押宝 JAX，其中都有哪些值得关注的信息？","谷歌在tensorflow发展不顺利的情况下，通过jax来重新定义api标准、在新赛道上（AI+科学计算）构建全新的能力是一种好的尝试。","2022-06-21","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/06/22/1ca38a1770274f37a054122f9e3ed745.png","technology-blogs","大V博文",{"type":15,"children":16,"toc":172},"root",[17,25,34,49,54,59,64,69,74,79,84,89,96,105,115,131,141,157],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"谷歌抛弃-tensorflow押宝-jax其中都有哪些值得关注的信息",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":18,"tag":30,"props":31,"children":33},"img",{"alt":7,"src":32},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/06/22/7535257effcf4ceb8a0401823f9bcea9.gif",[],{"type":18,"tag":26,"props":35,"children":36},{},[37,39,47],{"type":24,"value":38},"作者：金雪锋 链接：",{"type":18,"tag":40,"props":41,"children":45},"a",{"href":42,"rel":43},"https://www.zhihu.com/question/537823429/answer/2537694584",[44],"nofollow",[46],{"type":24,"value":42},{"type":24,"value":48}," 现在说谷歌抛弃TensorFlow，押宝JAX，估计还早，有点标题党。从使用者的角度看，当前除了谷歌外，其他公司使用JAX的感觉并不多，从Papers with code网站的统计看，论文的占比还比较少。但是JAX这个技术方向是很有意思的，这里代表了几个趋势：",{"type":18,"tag":26,"props":50,"children":51},{},[52],{"type":24,"value":53},"1、从业务上看，JAX的驱动力是AI+科学计算，和传统的CV/NLP相比，算是一个新赛道，从框架角度看，核心的能力是NN计算和数值计算的融合。",{"type":18,"tag":26,"props":55,"children":56},{},[57],{"type":24,"value":58},"金雪锋：AI框架的演进趋势和MindSpore的构想",{"type":18,"tag":26,"props":60,"children":61},{},[62],{"type":24,"value":63},"2、从接口看，兼容numpy、scipy等python原生的数据科学的接口，在此基础上扩展分布式（pmap/pjit）、向量化（vmap）、grad（各种高级微分）、JIT（硬件加速），既能兼容现有的Python生态，又在此基础上增加可微分能力，同时通过各种机制实现扩展性和性能加速。",{"type":18,"tag":26,"props":65,"children":66},{},[67],{"type":24,"value":68},"3、从编程风格上，全面采用函数式的风格，主要体现在无副作用、lamda闭包等；特别是闭包的特性可以使得各种功能可以灵活组合，比如：pmap（vmap（func））-pmap和vmap组合。另外，需要强调的是微分机制，这个与Pytorch的差异就更大一点，Jax的微分机制是函数式的自动微分机制，而Pytorch的是以tensor为中心的机制，我个人感觉函数式的其实更加符合算法的直观，下面有一个以前的简单分析：",{"type":18,"tag":26,"props":70,"children":71},{},[72],{"type":24,"value":73},"金雪锋：聊一聊AI框架前端",{"type":18,"tag":26,"props":75,"children":76},{},[77],{"type":24,"value":78},"4、实际上函数式这一套，pytorch也在跟随，但是个人觉得，与pytorch原来的设计理念有冲突。",{"type":18,"tag":26,"props":80,"children":81},{},[82],{"type":24,"value":83},"金雪锋：Pytorch 1.11看上去很美，但是.....",{"type":18,"tag":26,"props":85,"children":86},{},[87],{"type":24,"value":88},"不过，JAX目前看，并不算特别成熟，主要表现在，从传统的深度学习的角度看，nn的表达完全使用纯函数并不方便，所以后来jax也发展了很多nn封装的库，比如haiku等；另外，jax在分布式方面做的还不够成熟。总的来说，个人感觉，谷歌在tensorflow发展不顺利的情况下，通过jax来重新定义api标准、在新赛道上（AI+科学计算）构建全新的能力是一种好的尝试。",{"type":18,"tag":26,"props":90,"children":91},{},[92],{"type":18,"tag":30,"props":93,"children":95},{"alt":7,"src":94},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/06/22/140f840f8e8248fb9baa4c80bb5c0f80.jpg",[],{"type":18,"tag":26,"props":97,"children":98},{},[99],{"type":18,"tag":100,"props":101,"children":102},"strong",{},[103],{"type":24,"value":104},"MindSpore官方资料",{"type":18,"tag":26,"props":106,"children":107},{},[108,113],{"type":18,"tag":100,"props":109,"children":110},{},[111],{"type":24,"value":112},"官方QQ群",{"type":24,"value":114}," : 486831414",{"type":18,"tag":26,"props":116,"children":117},{},[118,123,125],{"type":18,"tag":100,"props":119,"children":120},{},[121],{"type":24,"value":122},"官网",{"type":24,"value":124},"：",{"type":18,"tag":40,"props":126,"children":129},{"href":127,"rel":128},"https://www.mindspore.cn/",[44],[130],{"type":24,"value":127},{"type":18,"tag":26,"props":132,"children":133},{},[134,139],{"type":18,"tag":100,"props":135,"children":136},{},[137],{"type":24,"value":138},"Gitee",{"type":24,"value":140}," : https : //gitee.com/mindspore/mindspore",{"type":18,"tag":26,"props":142,"children":143},{},[144,149,151],{"type":18,"tag":100,"props":145,"children":146},{},[147],{"type":24,"value":148},"GitHub",{"type":24,"value":150}," : ",{"type":18,"tag":40,"props":152,"children":155},{"href":153,"rel":154},"https://github.com/mindspore-ai/mindspore",[44],[156],{"type":24,"value":153},{"type":18,"tag":26,"props":158,"children":159},{},[160,165,166],{"type":18,"tag":100,"props":161,"children":162},{},[163],{"type":24,"value":164},"论坛",{"type":24,"value":124},{"type":18,"tag":40,"props":167,"children":170},{"href":168,"rel":169},"https://bbs.huaweicloud.com/forum/forum-1076-1.html",[44],[171],{"type":24,"value":168},{"title":7,"searchDepth":173,"depth":173,"links":174},4,[],"markdown","content:technology-blogs:zh:1562.md","content","technology-blogs/zh/1562.md","technology-blogs/zh/1562","md",1776506113423]