[{"data":1,"prerenderedAt":294},["ShallowReactive",2],{"content-query-vdWtTp12h2":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":288,"_id":289,"_source":290,"_file":291,"_stem":292,"_extension":293},"/news/zh/2547","zh",false,"","【昇思MindSpore技术公开课】第七节Prompt Tuning课程回顾来啦！","在上周六（6月3日）的昇思MindSpore技术功课中，我们向大家介绍了继 Pre-train + Fine-tune 后另一种新的微调范式—— Prompt Tuning，过去“通过调整模型以适配下游任务”的思路也由此转变为“通过调整下游任务描述以适配模型”。在课程中我们学习了Prompt Learning的全流程，以及对于Prompt的不同设计。接下来我们对课程进行简单回顾，迎接下一节公开课的进一步深入。","2023-06-07","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/06/08/3c3b546625284428bf010cf056ba506a.png","news",{"type":14,"children":15,"toc":285},"root",[16,24,29,38,46,51,61,70,78,83,88,93,98,103,108,116,121,126,131,138,143,150,155,163,168,173,197,204,209,214,221,229,237,242,247,255,260,278],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"昇思mindspore技术公开课第七节prompt-tuning课程回顾来啦",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":9},{"type":17,"tag":25,"props":30,"children":31},{},[32],{"type":17,"tag":33,"props":34,"children":35},"strong",{},[36],{"type":23,"value":37},"一、",{"type":17,"tag":25,"props":39,"children":40},{},[41],{"type":17,"tag":33,"props":42,"children":43},{},[44],{"type":23,"value":45},"课程回顾",{"type":17,"tag":25,"props":47,"children":48},{},[49],{"type":23,"value":50},"From Fine-tuning to Prompt Learning：",{"type":17,"tag":52,"props":53,"children":54},"ul",{},[55],{"type":17,"tag":56,"props":57,"children":58},"li",{},[59],{"type":23,"value":60},"NLP发展过程中的四个范式",{"type":17,"tag":25,"props":62,"children":63},{},[64],{"type":17,"tag":65,"props":66,"children":69},"img",{"alt":67,"src":68},"image.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230608030600.10337838809666680696032901356847:50540607070128:2400:494FF45DCEACD03F0892B4560752C6272D0B84F50B6F20D1BE9900B27C60F85A.png",[],{"type":17,"tag":52,"props":71,"children":72},{},[73],{"type":17,"tag":56,"props":74,"children":75},{},[76],{"type":23,"value":77},"Pre-train + Fine-tune回顾：",{"type":17,"tag":25,"props":79,"children":80},{},[81],{"type":23,"value":82},"1、BERT：Masked LM + NSP，注重文本理解",{"type":17,"tag":25,"props":84,"children":85},{},[86],{"type":23,"value":87},"2、GPT：auto-regressive model，注重文本生成",{"type":17,"tag":25,"props":89,"children":90},{},[91],{"type":23,"value":92},"3、Pre-train + Fine-tune面临的问题",{"type":17,"tag":25,"props":94,"children":95},{},[96],{"type":23,"value":97},"1）Pre-train 与 Fine-tune 间的语义差异",{"type":17,"tag":25,"props":99,"children":100},{},[101],{"type":23,"value":102},"2）由少样本引发的过拟合",{"type":17,"tag":25,"props":104,"children":105},{},[106],{"type":23,"value":107},"3） Fine-tune及模型部署成本较大",{"type":17,"tag":52,"props":109,"children":110},{},[111],{"type":17,"tag":56,"props":112,"children":113},{},[114],{"type":23,"value":115},"Pre-train, Prompt, Predict",{"type":17,"tag":25,"props":117,"children":118},{},[119],{"type":23,"value":120},"1、Prompting定义",{"type":17,"tag":25,"props":122,"children":123},{},[124],{"type":23,"value":125},"2、Workflow of Prompting",{"type":17,"tag":25,"props":127,"children":128},{},[129],{"type":23,"value":130},"1）在Prompting中最关键的两步便是对Prompt Template与Verbalizer的设计",{"type":17,"tag":25,"props":132,"children":133},{},[134],{"type":17,"tag":65,"props":135,"children":137},{"alt":67,"src":136},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230608030621.04020464281092317219086231186905:50540607070128:2400:FC6C9294BB8F99CD2D1AC2927325CBA0B3495D62A5843955CB309ACB90B82805.png",[],{"type":17,"tag":25,"props":139,"children":140},{},[141],{"type":23,"value":142},"2） Prompt分类",{"type":17,"tag":25,"props":144,"children":145},{},[146],{"type":17,"tag":65,"props":147,"children":149},{"alt":67,"src":148},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230608030639.50115656518211419326753100997099:50540607070128:2400:75D6F5B3451D1B07C338033C5555FFEB6A99E9ED192C87FA671A2ED910DEFF1D.png",[],{"type":17,"tag":25,"props":151,"children":152},{},[153],{"type":23,"value":154},"Prompt Template Engineering：",{"type":17,"tag":52,"props":156,"children":157},{},[158],{"type":17,"tag":56,"props":159,"children":160},{},[161],{"type":23,"value":162},"Prompt shape",{"type":17,"tag":25,"props":164,"children":165},{},[166],{"type":23,"value":167},"1、Cloze template",{"type":17,"tag":25,"props":169,"children":170},{},[171],{"type":23,"value":172},"2、Prefix template",{"type":17,"tag":52,"props":174,"children":175},{},[176,187],{"type":17,"tag":56,"props":177,"children":178},{},[179,181,185],{"type":23,"value":180},"Manual template learning",{"type":17,"tag":182,"props":183,"children":184},"br",{},[],{"type":23,"value":186},"模型并行是算子层面的并行，它利用某些算子的特性将算子拆分到多个设备上进行计算。",{"type":17,"tag":56,"props":188,"children":189},{},[190,192,195],{"type":23,"value":191},"Automated template learning",{"type":17,"tag":182,"props":193,"children":194},{},[],{"type":23,"value":196},"1、AutoPrompt：gradient based prompt search in discrete space",{"type":17,"tag":25,"props":198,"children":199},{},[200],{"type":17,"tag":65,"props":201,"children":203},{"alt":67,"src":202},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230608030701.56441883027169823238488599836441:50540607070128:2400:A9C11893665DB3124A7CF5B524594AED9F3809A6D92AEE4E71AD3512D910DFDE.png",[],{"type":17,"tag":25,"props":205,"children":206},{},[207],{"type":23,"value":208},"2、P-Tuning",{"type":17,"tag":25,"props":210,"children":211},{},[212],{"type":23,"value":213},"3、P-Tuning v2",{"type":17,"tag":25,"props":215,"children":216},{},[217],{"type":17,"tag":65,"props":218,"children":220},{"alt":67,"src":219},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230608030720.42757024016028214070765882698872:50540607070128:2400:386D41F27B28DA106E3B2FE9C5D13C9F56DBB403C5F3F51D5BE75FB55ADCB5ED.png",[],{"type":17,"tag":25,"props":222,"children":223},{},[224],{"type":17,"tag":33,"props":225,"children":226},{},[227],{"type":23,"value":228},"二、",{"type":17,"tag":25,"props":230,"children":231},{},[232],{"type":17,"tag":33,"props":233,"children":234},{},[235],{"type":23,"value":236},"下节课预告",{"type":17,"tag":25,"props":238,"children":239},{},[240],{"type":23,"value":241},"本周六（6月10日）我们邀请到中国科学院自动化研究所紫东.太初大模型研究中心常务副主任，武汉人工智能研究院院长王金桥老师，以及武汉人工智能研究院语音算法工程师唐帅老师，分别进行紫东.太初多模态大模型和语音识别的讲解。本周直播的课程内容有：",{"type":17,"tag":25,"props":243,"children":244},{},[245],{"type":23,"value":246},"紫东.太初多模态大模型 —— 王金桥：",{"type":17,"tag":52,"props":248,"children":249},{},[250],{"type":17,"tag":56,"props":251,"children":252},{},[253],{"type":23,"value":254},"紫东.太初多模态大模型2.0功能及应用介绍",{"type":17,"tag":25,"props":256,"children":257},{},[258],{"type":23,"value":259},"语音识别 —— 唐帅：",{"type":17,"tag":52,"props":261,"children":262},{},[263,268,273],{"type":17,"tag":56,"props":264,"children":265},{},[266],{"type":23,"value":267},"语音识别",{"type":17,"tag":56,"props":269,"children":270},{},[271],{"type":23,"value":272},"语音识别的系统框架",{"type":17,"tag":56,"props":274,"children":275},{},[276],{"type":23,"value":277},"语音识别现状及挑战",{"type":17,"tag":25,"props":279,"children":280},{},[281],{"type":17,"tag":65,"props":282,"children":284},{"alt":67,"src":283},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20230608030753.83891559257884029093815443044232:50540607070128:2400:CBFD92FF2C1D43EDDA53B99945885F0EC8BB826F75C4636DB67AC5B83232FA93.png",[],{"title":7,"searchDepth":286,"depth":286,"links":287},4,[],"markdown","content:news:zh:2547.md","content","news/zh/2547.md","news/zh/2547","md",1776506066898]