liheyoung / shrinkmatch Goto Github PK

I've read your paper very carefully and it's very inspiring, but I'm confused about one thing, for uncertain samples, each time you choose the class with the highest category probability and the remaining low-probability classes, which makes the network focus on the Top-1 classes only. If the Top-1 classes are wrongly predicted at the beginning, wouldn't the model be more severely biased when it is trained later on? Looking forward to hearing from you! Thank you!

Hi,

Thank you for releasing the code for this state of the art method. I am working on a comparative study across various Semi-Supervised learning methods, for which purpose availability of pre-trained ImageNet models for your method would ensure best possible evaluation.

Would it be possible for you to release your ImageNet models ?

Thanks.