Git Product home page Git Product logo

💡 [REQUEST] - <title>微调过后没有效果,有没有人知道,多少条数据会有效果 about qwen-vl HOT 8 OPEN

xuyiming010912 avatar xuyiming010912 commented on August 26, 2024
💡 [REQUEST] - 微调过后没有效果,有没有人知道,多少条数据会有效果<p>from qwen-vl.</p></section> </section> </article> <ins class="adsbygoogle" style="display:block" data-ad-client="ca-pub-7917632214101949" data-ad-slot="6627871389" data-ad-format="auto" data-full-width-responsive="true"></ins> <script> (adsbygoogle = window.adsbygoogle || []).push({}); </script> <article> <h2 class="h2">Comments (8)</h2> <section class="issue-comment"> <section id="2132586965" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/58521200?s=40&v=4" alt="liuhuan-gl avatar" /> <a class="issue-username" href="/liuhuan-gl">liuhuan-gl</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><p dir="auto">Me too</p><p>from qwen-vl.</p></section> </section> <section class="issue-comment"> <section id="2138612692" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/43362567?s=30&u=8c8025c017b3763d70cf8b8dbb9dd089dff830b4&v=4" alt="KDD2018 avatar" /> <a class="issue-username" href="/KDD2018">KDD2018</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><p dir="auto">我在2400+对图文定位数据集上做基于Lora的微调,效果很差,完全找不到图片目标和文本得对应关系,我也试着调整--fix-vit参数,但也没用,效果依旧很差。有大佬知道如何应对吗?</p><p>from qwen-vl.</p></section> </section> <section class="issue-comment"> <section id="2164672625" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/37207093?s=30&u=c0e1b559f2f4872bbf537263f923a542acbd1cfe&v=4" alt="elesun2018 avatar" /> <a class="issue-username" href="/elesun2018">elesun2018</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><p dir="auto">预训练模型用的哪个,应该是vl-chat<br> 效果,数据量 跟任务难度又有关系。<br> 主要还是看loss下降情况进行分析。</p> <p dir="auto">个人见解</p><p>from qwen-vl.</p></section> </section> <section class="issue-comment"> <section id="2164689637" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/148192069?v=4" alt="xuyiming010912 avatar" /> <a class="issue-username" href="/xuyiming010912">xuyiming010912</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><div class="email-fragment">找到原因了,训练时候用错模型了,用的量化的,合并的时候跟chat的合并的,导致于一系列的错误,但是检测框精度不是很高,泛化能力不是很大</div> <span class="email-hidden-toggle"><a href="#">…</a></span><div class="email-hidden-reply"> <div class="email-signature-reply">---- 回复的原邮件 ---- | 发件人 | ***@***.***> | | 日期 | 2024年06月13日 14:50 | | 收件人 | ***@***.***> | | 抄送至 | ***@***.***>***@***.***> | | 主题 | Re: [QwenLM/Qwen-VL] 💡 [REQUEST] - <title>微调过后没有效果,有没有人知道,多少条数据会有效果 (Issue <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="2314276637" data-permission-text="Title is private" data-url="https://github.com/QwenLM/Qwen-VL/issues/396" href="https://github.com/QwenLM/Qwen-VL/issues/396">#396</a>) | 预训练模型用的哪个,应该是vl-chat 效果,数据量 跟任务难度又有关系。 主要还是看loss下降情况进行分析。 个人见解 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***></div> </div><p>from qwen-vl.</p></section> </section> <ins class="adsbygoogle" style="display:block" data-ad-client="ca-pub-7917632214101949" data-ad-slot="6627871389" data-ad-format="auto" data-full-width-responsive="true"></ins> <script> (adsbygoogle = window.adsbygoogle || []).push({}); </script> <section class="issue-comment"> <section id="2164710985" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/37207093?s=30&u=c0e1b559f2f4872bbf537263f923a542acbd1cfe&v=4" alt="elesun2018 avatar" /> <a class="issue-username" href="/elesun2018">elesun2018</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><p dir="auto">我也正在摸索,可交流Q 294813364</p><p>from qwen-vl.</p></section> </section> <section class="issue-comment"> <section id="2164719337" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/148192069?v=4" alt="xuyiming010912 avatar" /> <a class="issue-username" href="/xuyiming010912">xuyiming010912</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><div class="email-fragment">有微信吗?QQ不常用</div> <span class="email-hidden-toggle"><a href="#">…</a></span><div class="email-hidden-reply"> <div class="email-signature-reply">---- 回复的原邮件 ---- | 发件人 | ***@***.***> | | 日期 | 2024年06月13日 14:58 | | 收件人 | ***@***.***> | | 抄送至 | ***@***.***>***@***.***> | | 主题 | Re: [QwenLM/Qwen-VL] 💡 [REQUEST] - <title>微调过后没有效果,有没有人知道,多少条数据会有效果 (Issue <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="2314276637" data-permission-text="Title is private" data-url="https://github.com/QwenLM/Qwen-VL/issues/396" href="https://github.com/QwenLM/Qwen-VL/issues/396">#396</a>) | 我也正在摸索,可交流Q 294813364 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***></div> </div><p>from qwen-vl.</p></section> </section> <section class="issue-comment"> <section id="2164726872" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/37207093?s=30&u=c0e1b559f2f4872bbf537263f923a542acbd1cfe&v=4" alt="elesun2018 avatar" /> <a class="issue-username" href="/elesun2018">elesun2018</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><p dir="auto">dcsun001</p><p>from qwen-vl.</p></section> </section> <section class="issue-comment"> <section id="2164734145" class="issue-head"> <img class="issue-avatar" src="https://avatars.githubusercontent.com/u/148192069?v=4" alt="xuyiming010912 avatar" /> <a class="issue-username" href="/xuyiming010912">xuyiming010912</a> <span class="issue-time"> commented on August 26, 2024 </span> </section> <section class="markdown markdown-js p-5"><div class="email-fragment">好的,备注是徐一鸣</div> <span class="email-hidden-toggle"><a href="#">…</a></span><div class="email-hidden-reply"> <div class="email-signature-reply">---- 回复的原邮件 ---- | 发件人 | ***@***.***> | | 日期 | 2024年06月13日 15:03 | | 收件人 | ***@***.***> | | 抄送至 | ***@***.***>***@***.***> | | 主题 | Re: [QwenLM/Qwen-VL] 💡 [REQUEST] - <title>微调过后没有效果,有没有人知道,多少条数据会有效果 (Issue <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="2314276637" data-permission-text="Title is private" data-url="https://github.com/QwenLM/Qwen-VL/issues/396" href="https://github.com/QwenLM/Qwen-VL/issues/396">#396</a>) | dcsun001 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***></div> </div><p>from qwen-vl.</p></section> </section> </article> <section> <h2 class="h2">Related Issues (20)</h2> <div class="issue"> <ul> <li> <a href="/qwenlm/qwen-vl/issues/415">关于模型融合</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 16</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/416">[BUG] <qwen-vl api 在阿里云ecs 上调用出现 网络连接错误></a> </li> <li> <a href="/qwenlm/qwen-vl/issues/417">关于 chat模型 和 base模型的微调</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/418">多卡推理错误[BUG] <title></a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 1</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/419">如何使用langchain使用qwen-vl_max 模型</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/420">💡 [REQUEST] - 可以直接使用Qwen-VL-plus或者Qwen-VL-max吗?</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/421">About Qwen2-vl model</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/422">qwen2-VL 图文多模态模型有发布计划吗?</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 1</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/423">是否有jetson的部署方案推荐</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/424">有Qwen2-VL-Instruct 图文多模态模型的发布计划吗?</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 1</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/425">💡 [REQUEST] - <请问gptq量化相关的工程代码可否开源?></a> </li> <li> <a href="/qwenlm/qwen-vl/issues/426">[BUG] <推理阶段,模型forward方法的visual分支并未进行视觉编码></a> </li> <li> <a href="/qwenlm/qwen-vl/issues/427">AssertionError: Only Support Self-Attention Currently</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 2</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/428">[Help] Qwen-VL model.generate方法如何输出output_attention</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 1</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/429">How to get a better result with finetune(如何通过finetune得到一个较好的结果)</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 5</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/430">[BUG] <title>api 请求报错</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/431">[BUG] <调用qwen_vl_max>接口,传入图片后报下载图片错误</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/432">训练数据中对一张图片如果存在100轮QA,应如何制作训练数据集</a> </li> <li> <a href="/qwenlm/qwen-vl/issues/433">[BUG] <title> Qwen-VL-Chat-Int4 load进入infer时提示有很多weights没有使用</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 1</span> </li> <li> <a href="/qwenlm/qwen-vl/issues/434">拉了一个多模态大模型技术交流群,大家可以加入进来进行技术交流</a> <span class="text-red-600 text-xs font-normal py-0.5 px-1 border border-red-600 rounded-md">HOT 1</span> </li> </ul> </div> </section> </main> <section id="more" class="flex-none w-full md:w-60 text-gray-600 bg-gray-50 px-5 md:px-3 rounded-md dark-color"> <div class="w-full md:w-60 h-0.5"></div> <section> <!-- recommend projects --> <h2 class="h2 py-3.5">Recommend Projects</h2> <ul> <li class="mb-4"> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/facebook/react"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://raw.githubusercontent.com/facebook/create-react-app/master/packages/cra-template/template/public/logo192.png" alt="React photo" /> React </a> </h3> <p class="article-more pt-1">A declarative, efficient, and flexible JavaScript library for building user interfaces.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/vuejs/vue"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://camo.githubusercontent.com/c8f91d18976e27123643a926a2588b8d931a0292fd0b6532c3155379e8591629/68747470733a2f2f7675656a732e6f72672f696d616765732f6c6f676f2e706e67" alt="Vue.js photo" /> Vue.js </a> </h3> <p class="article-more pt-1">🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/microsoft/TypeScript"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://www.typescriptlang.org/favicon-32x32.png" alt="Typescript photo" /> Typescript </a> </h3> <p class="article-more pt-1">TypeScript is a superset of JavaScript that compiles to clean JavaScript output.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/tensorflow/tensorflow"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://camo.githubusercontent.com/c04e16c05de80dadbdc990884672fc941fdcbbfbb02b31dd48c248d010861426/68747470733a2f2f7777772e74656e736f72666c6f772e6f72672f696d616765732f74665f6c6f676f5f736f6369616c2e706e67" alt="TensorFlow photo" /> TensorFlow </a> </h3> <p class="article-more pt-1">An Open Source Machine Learning Framework for Everyone</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/django/django"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars2.githubusercontent.com/u/27804?s=200&v=4" alt="Django photo" /> Django </a> </h3> <p class="article-more pt-1">The Web framework for perfectionists with deadlines.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/laravel/laravel"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://laravel.com/img/logomark.min.svg" alt="Laravel photo" /> Laravel </a> </h3> <p class="article-more pt-1">A PHP framework for web artisans</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/d3/d3"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://camo.githubusercontent.com/586ccf0aad9684edc821658cee04146cf36d1f1d5ec904bbefd72728909ccb2e/68747470733a2f2f64336a732e6f72672f6c6f676f2e737667" alt="D3 photo" /> D3 </a> </h3> <p class="article-more pt-1">Bring data to life with SVG, Canvas and HTML. 📊📈🎉</p> </article> </li> <li> <div> <ins class="adsbygoogle" style="display:block" data-ad-client="ca-pub-7917632214101949" data-ad-slot="6627871389" data-ad-format="auto" data-full-width-responsive="true"></ins> <script> (adsbygoogle = window.adsbygoogle || []).push({}); </script> </div> </li> </ul> </section> <section> <!-- recommend topics --> <h2 class="h2 py-3.5">Recommend Topics</h2> <ul> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/topic/javascript"> javascript </a> </h3> <p class="article-more pt-1">JavaScript (JS) is a lightweight interpreted programming language with first-class functions.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/topic/web"> web </a> </h3> <p class="article-more pt-1">Some thing interesting about web. New door for the world.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/topic/server"> server </a> </h3> <p class="article-more pt-1">A server is a program made to process requests and deliver data to clients.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/topic/machine-learning"> Machine learning </a> </h3> <p class="article-more pt-1">Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/topic/visualization"> Visualization </a> </h3> <p class="article-more pt-1">Some thing interesting about visualization, use data art</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/topic/game"> Game </a> </h3> <p class="article-more pt-1">Some thing interesting about game, make everyone happy.</p> </article> </li> <li> <ins class="adsbygoogle" style="display:block" data-ad-client="ca-pub-7917632214101949" data-ad-slot="6627871389" data-ad-format="auto" data-full-width-responsive="true"></ins> <script> (adsbygoogle = window.adsbygoogle || []).push({}); </script> </li> </ul> </section> <section> <!-- recommend users --> <h2 class="h2 py-3.5">Recommend Org</h2> <ul> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/facebook"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars.githubusercontent.com/u/69631?v=4" alt="Facebook photo" /> Facebook </a> </h3> <p class="article-more pt-1">We are working to build community through open source technology. NB: members must have two-factor auth.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/microsoft"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars.githubusercontent.com/u/6154722?v=4" alt="Microsoft photo" /> Microsoft </a> </h3> <p class="article-more pt-1">Open source projects and samples from Microsoft.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/google"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars.githubusercontent.com/u/1342004?v=4" alt="Google photo" /> Google </a> </h3> <p class="article-more pt-1">Google ❤️ Open Source for everyone.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/alibaba"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars.githubusercontent.com/u/1961952?v=4" alt="Alibaba photo" /> Alibaba </a> </h3> <p class="article-more pt-1">Alibaba Open Source for everyone</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/d3"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars.githubusercontent.com/u/1562726?v=4" alt="D3 photo" /> D3 </a> </h3> <p class="article-more pt-1">Data-Driven Documents codes.</p> </article> </li> <li> <article class="small-box"> <h3 class="article-title"> <a class="block break-all" href="/tencent"> <img loading="lazy" class="inline-block w-6 h-6 rounded-md border border-white" width="24" height="24" src="https://avatars.githubusercontent.com/u/18461506?v=4" alt="Tencent photo" /> Tencent </a> </h3> <p class="article-more pt-1">China tencent open source team.</p> </article> </li> <li> <ins class="adsbygoogle" style="display:block" data-ad-client="ca-pub-7917632214101949" data-ad-slot="6627871389" data-ad-format="auto" data-full-width-responsive="true"></ins> <script> (adsbygoogle = window.adsbygoogle || []).push({}); </script> </li> </ul> </section> </section> </div> </div> <!-- footer --> <footer class="sizeing text-xs text-center p-5"> <div>Friends: <a class="hover:underline" target="_blank" href="https://www.chanpinqingbaoju.com">ProductDiscover</a> </div> Copyright © 2024 Git Product <!-- & <span class="block md:inline">Data Power by github.com</span> --> ❤️ <a class="hover:underline block md:inline" href="mailto:cs.victor.edison@gmail.com">Mail to me</a> </footer> </body> </html>