distill_visual_priors

This is the 2nd place solution of ECCV 2020 workshop VIPriors Image Classification Challenge.

The two phases of our proposed method. The first phase is to construct a useful visual prior with self-supervised contrastive learning, and the second phase is to perform self-distillation on the pre-trained checkpoint. The student model is trained with a distillation loss and a classification loss, while the teacher model is frozen.

Usage

Our solution presents a two-phase pipeline, and we only use the provided subset of ImageNet, no external data or checkpoint is used in our solution.

Phase-1

Self-supervised pretraining.

Please follow the instructions in the moco folder.

Phase-2

Self-distillation and classification finetuning.

cd sup_train_distill
python3 train_selfsup.py --data_path /path/to/data/ --net_type self_sup_r50 --input-res 448 --pretrained /path/to/unsupervise_pretrained_checkpoint --save_path /path/to/save --batch_size 256 --autoaug --label_smooth

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follow.

@inproceedings{zhao2020distilling,
  title={Distilling visual priors from self-supervised learning},
  author={Zhao, Bingchen and Wen, Xin},
  booktitle={European Conference on Computer Vision},
  pages={422--429},
  year={2020},
  organization={Springer}
}

Contact

Bingchen Zhao: zhaobc.gm@gmail.com