search

microsoft / SoftTeacher

Semi-Supervised Learning, Object Detection, ICCV2021
MIT License
904 stars 123 forks source link

What is the effect of using no augmentation instead of weak augmentation for teacher? #135

Closed lhphanto closed 2 years ago

lhphanto commented 2 years ago

Hi,

Just wonder if it is ok to have no data augmentation for teacher model (instead of weak augmentation)?

Thanks!

jackhu-bme commented 2 years ago

Well, I guess very weak data augmentation like random flip is used as regularization and is essential for the generalization of model, theoretically. Maybe mAP will drop a little, you could try this as a practice. Actually, I don't think this setting change makes much sense for this situation.

lhphanto commented 2 years ago

Thanks, Jack! So I thought the reason why we have different augmentation for teacher and student is to do some kind of consistency matching? If so, I was wondering why we can't do all the augmentation in the student. And leave teacher alone, which is used for generating labels, and is like a eval/inference job. And for eval jobs, we don't necessarily need data augmentation, I think?

On Wed, Dec 29, 2021, 5:56 AM Jack Hu @.***> wrote:

Well, I guess very weak data augmentation like random flip is used as regularization and is essential for the generalization of model, theoretically. Maybe mAP will drop a little, you could try this as a practice. Actually, I don't think this setting change makes much sense for this situation.

— Reply to this email directly, view it on GitHub https://github.com/microsoft/SoftTeacher/issues/135#issuecomment-1002606164, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABISA2DH2KYE4L42BY4FQB3UTMHP7ANCNFSM5K5MUBUQ . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

You are receiving this because you authored the thread.Message ID: @.***>

jackhu-bme commented 2 years ago

Just like augment test, when testing, weak augment improves mAP. Same reason here, weak augment improves the quality of psuedo labels, I guess.

jackhu-bme commented 2 years ago

Like, here. https://machinelearningmastery.com/how-to-use-test-time-augmentation-to-improve-model-performance-for-image-classification/