search

YisenWang / symmetric_cross_entropy_for_noisy_labels

Code for ICCV2019 "Symmetric Cross Entropy for Robust Learning with Noisy Labels"
https://arxiv.org/abs/1908.06112
169 stars 32 forks source link

Asymmetric Noise- #10

Open jaideep11061982 opened 3 years ago

jaideep11061982 commented 3 years ago

Hi what are optimal values for alpha beta in acse my noise rate is around 0.2 and type asymmetric

YisenWang commented 3 years ago

Hi, which dataset do you use?

jaideep11061982 commented 3 years ago

Its a customer dataset 21k images, best accuracy we get 89 on the set

On Thu, 18 Feb 2021, 16:27 YisenWang, notifications@github.com wrote:

Hi, which dataset do you use?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels/issues/10#issuecomment-781260275, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJDFNZCOLAXHAOKTBABNPTLS7TXANANCNFSM4XX2OYHA .

YisenWang commented 3 years ago

I think you could tune the alpha and beta using grid search to adaptively fit your customer dataset.

On 19/02/2021, jaideep11061982 notifications@github.com wrote:

Its a customer dataset 21k images, best accuracy we get 89 on the set

On Thu, 18 Feb 2021, 16:27 YisenWang, notifications@github.com wrote:

Hi, which dataset do you use?

—

You are receiving this because you authored the thread.

Reply to this email directly, view it on GitHub

https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels/issues/10#issuecomment-781260275,

or unsubscribe

https://github.com/notifications/unsubscribe-auth/AJDFNZCOLAXHAOKTBABNPTLS7TXANANCNFSM4XX2OYHA

.

--

You are receiving this because you commented.

Reply to this email directly or view it on GitHub:

https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels/issues/10#issuecomment-781479025

-- Yisen Wang Department of Computer Science and Technology Tsinghua University Beijing, P.R.China(100084) Email: eewangyisen@gmail.com wangys14@mails.tsinghua.edu.cn

jaideep11061982 commented 3 years ago

Ok in general This loss helps prevent noise free label from getting wrong by not letting model overfit to noisy data or helps predicting images which have noisy labels

On Thu, 18 Feb 2021, 22:17 YisenWang, notifications@github.com wrote:

I think you could tune the alpha and beta using grid search to adaptively fit your customer dataset.

On 19/02/2021, jaideep11061982 notifications@github.com wrote:

Its a customer dataset 21k images, best accuracy we get 89 on the set

On Thu, 18 Feb 2021, 16:27 YisenWang, notifications@github.com wrote:

Hi, which dataset do you use?

—

You are receiving this because you authored the thread.

Reply to this email directly, view it on GitHub

< https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels/issues/10#issuecomment-781260275 ,

or unsubscribe

< https://github.com/notifications/unsubscribe-auth/AJDFNZCOLAXHAOKTBABNPTLS7TXANANCNFSM4XX2OYHA

.

--

You are receiving this because you commented.

Reply to this email directly or view it on GitHub:

https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels/issues/10#issuecomment-781479025

-- Yisen Wang Department of Computer Science and Technology Tsinghua University Beijing, P.R.China(100084) Email: eewangyisen@gmail.com wangys14@mails.tsinghua.edu.cn

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels/issues/10#issuecomment-781480780, or unsubscribe https://github.com/notifications/unsubscribe-auth/AJDFNZGY6GCVEOKT5UTY4LTS7VABDANCNFSM4XX2OYHA .