search

YU1ut / MixMatch-pytorch

Code for "MixMatch - A Holistic Approach to Semi-Supervised Learning"
MIT License
633 stars 129 forks source link

Question about the efficiency of parameter T #15

Open sanyouwu opened 5 years ago

sanyouwu commented 5 years ago

Thanks for your contribution. In the paper, the author did some ablation study. For verify the accuracy of your work, I repeat these configurations and get satisfying results. However, I find an amazing result, that is the accuracy under T = 1 ( Mixmatch without temperature sharpening) is larger than T = 0.5. ## I only change parameter T, and other parameters are the same as yours.

So, can I get an inference that temperature sharpening operate does not work? I am not confident about this since the author emphasizes its importance in the paper.@YU1ut

YU1ut commented 5 years ago

Thank you for reporting this. I will investigate this.