Question about burn-in steps

facebookresearch / unbiased-teacher

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection

https://arxiv.org/abs/2102.09480

MIT License

410 stars 82 forks source link

Question about burn-in steps #6

Closed Chrisfsj2051 closed 3 years ago

Chrisfsj2051 commented 3 years ago

Hi! It's exciting to see such work with remarkable improvements.

However, I have a question about the setting of the burn-in steps. In the paper, you mentioned that different step values are set for different data settings (e.g. 20k for COCO@10%). But it seems like that you set steps to 2000 in all configs, what is the reason?

Chrisfsj2051 commented 3 years ago

Besides, what is the value of burn-in steps that you used in VOC experiments?

ycliu93 commented 3 years ago

Hi @Chrisfsj2051,

We found out using 20k or 2k iterations for burn-in could converge to a similar mAP at the end of the training.
For the burn-in steps, we use 30k iterations.

Thanks!