search

TencentYoutuResearch / Classification-SemiCLS

Code for CVPR 2022 paper “Class-Aware Contrastive Semi-Supervised Learning”
Other
93 stars 13 forks source link

STL10 results of fixmatch not correct #3

Closed sailist closed 2 years ago

sailist commented 2 years ago

Thank you for your beautiful work! But there is a question: In the original fixmatch paper, the stl10 error rate is 5.17±0.63(94.93 accuracy), however the result on your paper is 65.38±0.42, are there any differences in your experiments?

KaiWU5 commented 2 years ago

Thanks for kindly mentioning. In original Fixmatch and MixMatch, WRN-37-2 is written to be used for STL-10. But the way to implement WRN-37-2 is not like WRN-28-2 because 37 cannot be divided by 2. In https://github.com/google-research/fixmatch/issues/34, David said they even got number of parameters wrong for WRN-37-2. So, to make experiments simple and consistent, we use Resnet-18 for STL10, the same in Comatch paper as mentioned in our paper.