Giter VIP home page Giter VIP logo

Comments (21)

Fight-hawk avatar Fight-hawk commented on May 23, 2024

看起来貌似teacher-model没有起到作用 是student-model自己在学习,是不是蒸馏损失的权重太小了?

from fgd.

renyu-renyu avatar renyu-renyu commented on May 23, 2024

I am researching this direction, can I exchange some questions about FGD with you?

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

@renyu-renyu ok,by the way, if you are Chinese, we can talk with native language.

from fgd.

renyu-renyu avatar renyu-renyu commented on May 23, 2024

@Fight-hawk 你好,可以一起交流吗

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

@renyu-renyu 可以啊 我最近也在搞知识蒸馏

from fgd.

renyu-renyu avatar renyu-renyu commented on May 23, 2024

@renyu-renyu 可以啊 我最近也在搞知识蒸馏

你有在其他算法实现吗

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

@renyu-renyu 其他目标检测算法么,目前只在yolox上进行了实验

from fgd.

renyu-renyu avatar renyu-renyu commented on May 23, 2024

@renyu-renyu 其他目标检测算法么,目前只在yolox上进行了实验

效果怎么样阿

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

训练150epoch 提升了AP0.5 提高了1.5 mmdet

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

@www516717402 我没复现精度,在自己的数据集上试了一下,有效果。作者一直拒绝提供完整的训练log,可能是加了一些trick吧

from fgd.

www516717402 avatar www516717402 commented on May 23, 2024

@Fight-hawk @renyu-renyu mmdet是在V2.15.1版本加入的YOLOX(当时很不稳定,而且只有tiny版本复现出了精度),而这个蒸馏用的是mmdet-V2.11版本。是他自己复现的YOLOX-m的baseline是45.9%???但是官方YOLOX-m的baseline是46.9%,不清楚作者用的是什么版本,还有如何复现的精度,这baseline精度差距1%去哪了。

from fgd.

www516717402 avatar www516717402 commented on May 23, 2024

@Fight-hawk 你自己的数据集尺度有coco那么大吗?还有AP0.5提高了1.5%,其它项有降低吗?一句题外话,蒸馏一般不适用YOLOX-L和YOLOX-Nano配合,Gap差距太大非常容易崩,你能有提高也是奇迹

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

不明白你在逼逼什么 蒸馏一般不适用YOLOX-L和YOLOX-Nano配合 关于你有实践过么 道听途说也能对别人逼逼赖赖?

from fgd.

www516717402 avatar www516717402 commented on May 23, 2024

@Fight-hawk
1)首先非常抱歉打扰你了。
2)蒸馏过人脸识别,大模型和超小模型有gap,很难训练而且效果不好。
3)仅仅是提出我的疑惑,没有任何别的意思。冒犯到你,再次抱歉。

from fgd.

yzd-v avatar yzd-v commented on May 23, 2024

@Fight-hawk @renyu-renyu mmdet是在V2.15.1版本加入的YOLOX(当时很不稳定,而且只有tiny版本复现出了精度),而这个蒸馏用的是mmdet-V2.11版本。是他自己复现的YOLOX-m的baseline是45.9%???但是官方YOLOX-m的baseline是46.9%,不清楚作者用的是什么版本,还有如何复现的精度,这baseline精度差距1%去哪了。

我用的mmdet==2.19做的yolox,mmdet的config跑出来就是45.9..........,对于yolox的修改是有yolox分支

from fgd.

yzd-v avatar yzd-v commented on May 23, 2024

@www516717402 我没复现精度,在自己的数据集上试了一下,有效果。作者一直拒绝提供完整的训练log,可能是加了一些trick吧

没有其他trick,用mmdet==2.19,再根据yolox分支进行修改一下就可以跑到结果的点

from fgd.

www516717402 avatar www516717402 commented on May 23, 2024

@Fight-hawk @renyu-renyu mmdet是在V2.15.1版本加入的YOLOX(当时很不稳定,而且只有tiny版本复现出了精度),而这个蒸馏用的是mmdet-V2.11版本。是他自己复现的YOLOX-m的baseline是45.9%???但是官方YOLOX-m的baseline是46.9%,不清楚作者用的是什么版本,还有如何复现的精度,这baseline精度差距1%去哪了。

我用的mmdet==2.19做的yolox,mmdet的config跑出来就是45.9..........,对于yolox的修改是有yolox分支

多谢解惑,yolox-m不达标可以提bug了。

from fgd.

www516717402 avatar www516717402 commented on May 23, 2024

@yzd-v 我跑了一下代码没问题(精度确实达不到官方)。感谢开源!

from fgd.

lixiangMindSpore avatar lixiangMindSpore commented on May 23, 2024

您好,您是怎么跑的?我看FGD/configs/distillers/fgd目录下没有相关yolox的配置?方便告知一下吗?多谢

from fgd.

Fight-hawk avatar Fight-hawk commented on May 23, 2024

您好,您是怎么跑的?我看FGD/configs/distillers/fgd目录下没有相关yolox的配置?方便告知一下吗?多谢

在yolox分支

from fgd.

yingwenmingzi123 avatar yingwenmingzi123 commented on May 23, 2024

您好,我想问一下你跑的那个版本的mmdet啊,mmdet==2.19的话会报sgd错误
ValueEPror: SGD: some parameters appear in more than one parameter group。方便告知一下这个是怎么解决的嘛。多谢

from fgd.

Related Issues (20)

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.