question

[nadongjin]

Hello ^^
Long time no see.
I have a question about the normalize of the image in your code.

normalize=lambda xx: (tf.cast(tf.image.decode_image(xx, channels=3), "float32") - 127.5) * 0.0078125

As above, you did 0.0078125 = 1/128.
But when I looked at the insightface's code, it was 1/128 for face detection and 1/127.5 for face recognition.
Normally, normalize -1<=x<=1. I wonder if there is any other reason for 1/128 instead of 1/127.5.
Thank you.

Ms1mv3 image cleaning is too painful.^^
I don’t' know the reason.
Why is there so much data that separated one class from the dataset?

[leondgarse]

0.0078125 is actually from the very early implementation of the mxnet insightface. IMO, the model training is not that sensitive for this 0.5 difference, either 127.5, 1 / 127.5 or 128, 1 / 128 or 127.5, 1 / 128, just keeping them same for both training and evaluation is OK.
Ya, manually cleaning the dataset is very brave! Haven't got time for my model cleaning...

[nadongjin]

I will send you a clean dataset at the end of this month.
I think this will only help me if you find a sota architecutre from Enjinier's point of view. ^^

[leondgarse]

Ya, it's still better using the raw dataset for comparing with others. Besides, I have finished some running use MagfaceLoss, and achieves IJBB 0.957449, IJBC 0.971212 using EfficienNetV2S + AdamW, and the training much stable than ArcFace at the beginning. May upload it soon, need some more tests.

[nadongjin]

I'm looking forward to it ^^

[leondgarse]

It's uploaded, with a higher result IJBB 0.958325, IJBC 0.971212. Enjoy.

[nadongjin]

You're a genius. :) I am still working on data clean work. It is questionable whether your current architecture and train techniques can be further improved. I think data improvement is the only way to improve your performance from now on. I will do my best to work hard and give it to you quickly. I wonder if the magface algorithm is definitely stable during learning. Once again, thank you for your efforts. 2022년 2월 24일 (목) 오후 4:37, leondgarse ***@***.***>님이 작성:

…

It's uploaded, with a higher result IJBB 0.958325, IJBC 0.971212. Enjoy. — Reply to this email directly, view it on GitHub <#81 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADRFGOCPVZBXFS3T2ZK3SXDU4XNZ5ANCNFSM5OPD5ITA> . 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: ***@***.*** com>

-- 유빈아빠

[leondgarse]

It's just some repeating works of parameters tuning. Lol
Ya, I don't know if its' reproducible. Just finally a recipe can applying AdamW for EfficientNetV2M, the previous AdamW + ArcFace actually most time not working well for V2M in my tests...