search

facebookresearch / unbiased-teacher

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection
https://arxiv.org/abs/2102.09480
MIT License
415 stars 83 forks source link

How to evaluate the performance of labeled data? #43

Closed wjn922 closed 3 years ago

wjn922 commented 3 years ago

Hi,

How to train the supervised model by only using 1% / 2% / 5% / 10% data?

Further question: How to train using the coco-additional (100% coco + unlabeled data)

ycliu93 commented 3 years ago

Since our burn-in stage is actually the vanilla supervised training. You could extend the burn-in stage iterations, and that is the supervised-only baseline with the labeled set only (1%-10%). And, don't forget to add the learning rate decay back for the supervised baseline for a fair comparison.