Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

关于MDD算法在Visda数据集上的复现 #48

Closed
dingning97 opened this issue May 5, 2021 · 6 comments
Closed

关于MDD算法在Visda数据集上的复现 #48

dingning97 opened this issue May 5, 2021 · 6 comments

Comments

@dingning97
Copy link

MDD原文中(Bridging Theory and Algorithm for Domain Adaptation),Table3写道算法在visda数据集上Acc能达到74%+,请问可以分享一下超参数设置吗?比如说Initial LR、lr_gamma、lr_decay。十分感谢,最近我也在进行DA的工作,但是利用该lib开源代码的default setting无法reproduce,这样不太敢直接引用数据。
十分感谢!!!!

@dingning97
Copy link
Author

贵组维护的MDD原repo(https://github.com/thuml/MDD/issues)似乎一直没有回复issues
Looking forward to reply and share !!!!! THX!

@JunguangJiang
Copy link
Collaborator

The following script should achieve ~74.9% on VisDA2017.
CUDA_VISIBLE_DEVICES=0 python mdd.py data/visda-2017 -d VisDA2017 -s Synthetic -t Real -a resnet50 --epochs 30
--bottleneck-dim 1024 --seed 0 --center-crop --per-class-eval -b 36 --log logs/mdd/VisDA2017_resnet50

@JunguangJiang
Copy link
Collaborator

default setting无法reproduce,这是什么意思呢?
你可以把你的训练脚本,以及实验结果列举一下吗?

@dingning97
Copy link
Author

image
这是我的训练结果,用的是args里面的default的值,不过我设置seed为1了,我觉得seed的值对训练结果还是影响挺大的。
而且我没加--center-crop这个参数,我觉得这也可能是我没复现的原因;麻烦问一下,在visda数据集上用center-crop的方式训练,是否有助于涨点呀?
我再train一个,明天早晨我再来回复你呀~~

@dingning97
Copy link
Author

image
这是我run的脚本CUDA_VISIBLE_DEVICES=0 python mdd.py data/visda-2017 -d VisDA2017 -s Synthetic -t Real -a resnet50 --epochs 100 -i 500 --bottleneck-dim 1024 -b 36 --log logs/mdd/VisDA2017 --seed 0 --center-crop
共训练5万个iter,best accuracy 74.21@iter27000,多跑几次应该差不多能跑出原文结果。

还训了一个没加--center-crop的模型,绿线,结果确实不好。

@JunguangJiang
Copy link
Collaborator

--center-crop对VisDA结果影响比较大,这也是MDD原实现中使用的方法

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants