GithubHelp home page GithubHelp logo

Comments (24)

Baijiong-Lin avatar Baijiong-Lin commented on August 28, 2024

已修复

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

哥们,还是有报错的,alpha可能是numpy也可能是tensor

from libmtl.

Baijiong-Lin avatar Baijiong-Lin commented on August 28, 2024

哪个配置下有报错,麻烦提供一下运行命令,谢谢

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

恩恩,好的,我明天在跑一下告诉你。我新发现一个nashmtl的问题你能指导我一下吗?我传入2任务的loss,数值分别是700多和600多,结果乘上nashmtl后,总的loss变成了0.5,有什么超参数可以调节的吗,这导致我模型收敛不了。

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

哪个配置下有报错,麻烦提供一下运行命令,谢谢

不好意思,是我看错了你的commit,修改错了,你的修改是ok的,已经没有报错了。另外,你能帮我看看我哦的bug吗,nashmtl算出来的weight_loss好小,是10-4级别的,是要调整什么超参数吗?谢谢

from libmtl.

Baijiong-Lin avatar Baijiong-Lin commented on August 28, 2024

这个问题可能需要问一下Nash-MTL的作者们

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

这个问题可能需要问一下Nash-MTL的作者们

问了。但感觉,梯度很大时候,nashmtl就会计算出很小的loss权重(10-4),反传回去时候网络参数的梯度就会非常小,整个网络根本迭代不动。

from libmtl.

Baijiong-Lin avatar Baijiong-Lin commented on August 28, 2024

我也没想到什么好方法

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

下有报错,麻烦提供一下运行命令,谢

这个问题可能需要问一下Nash-MTL的作者们

问了。但感觉,梯度很大时候,nashmtl就会计算出很小的loss权重(10-4),反传回去时候网络参数的梯度就会非常小,整个网络根本迭代不动。

请问,你这边用多任务算法的时候,有什么实际work的算法吗,我这边在实际使用的时候其实都会比单任务差很多

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

我这边主要是yolox架构的object detection跟自驾任务中的freespace,这两个任务,每个任务都有若干loss,但是用MTL来优化其实远比我随便调一下loss权重要差

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

@MingChaoXu
你都试了哪些方法

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

我尝试了pcgrad,uw,nash mtl,cagrad,graddrop,都不好使,不知道你这有没有work的经验?

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

@MingChaoXu
和nashmtl的作者聊的,可能要对输出做归一化,AvivNavon/nash-mtl#13
你的nash mtl没问题吗

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

我这边nash mtl可以收敛,就是精度变差了,你有尝试别的算法吗?效果怎么样?

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

我就是pcgrad,有一点点效果,试了nashmtl不行,现在在尝试gradnorm,我也是用的yolox的主干。感觉可能和你的数据分布相关

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

我的两个任务的数据并没有共同标注,就是每份数据只有一个任务的标注,不知道这个会不会影响,因为我看论文里的数据好像是多任务标注都有的?

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

这不是非同源数据吗?

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

我尝试了pcgrad,uw,nash mtl,cagrad,graddrop,都不好使,不知道你这有没有work的经验?

image

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

是的,可以说非同源数据吧,domain是一样的,只不过只有单个数据的标注,你的任务是两种标注都有是吗?

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

shide

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

谢谢作者,nashmtl这个方法我先不用了吧。。。。

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

是的,可以说非同源数据吧,domain是一样的,只不过只有单个数据的标注,你的任务是两种标注都有是吗?

可以看看这篇paper https://arxiv.org/pdf/2209.11379.pdf
另外,可以加下微信,讨论下吗,DayDayAmazing,备注mtl

from libmtl.

yushengjiexy avatar yushengjiexy commented on August 28, 2024

@MingChaoXu

from libmtl.

MingChaoXu avatar MingChaoXu commented on August 28, 2024

@yushengjiexy 好的~互相学习一下hhh,加你了

from libmtl.

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.