wvangansbeke / Unsupervised-Classification

SCAN: Learning to Classify Images without Labels, incl. SimCLR. [ECCV 2020]
https://arxiv.org/abs/2005.12320
Other
1.37k stars 268 forks source link

Overcoming uncertainity after scan phase by dynamically lowering accuracy threshold for self-labeling #128

Open TomasPlachy opened 2 years ago

TomasPlachy commented 2 years ago

Hello,

The first batch in the first epoch usually doesn't have confident enough samples to pass the 0.99 threshold for self-labeling. But if I lower the threshold to 0.9, it negatively effects the performance of the network (I have observed that most of the samples have over 0.99 probability score in the final epochs, but the accuracy is low).

Do you have any thoughts about progressively increasing the threshold for self-labeling? Or how would you tackle this issue?