VirFace

This is a PyTorch implementation of "VirFace: Enhancing Face Recognition via Unlabeled Shallow Data" (CVPR 2021).

Training stage

Training Stage	Datasets	args.method	pretrain model needed
Pre-train backbone	Label	pretrain	None
Pre-train generator	Label	generator	backbone, head
Training VirClass	Label & Unlabel	virclass	backbone, head
Training VirFace	Label & Unlabel	virface	backbone, head, generator

Backbone model

Arch	ResNet					User Arch
	18	34	50	101	152
args.arch	resnet18	resnet34	resnet50	resnet101	resnet152	usr

- How to use **user modified architecture**:

• edit config.py: import modified '.py' file, then modifiy "model_usr=None" to "model_usr=%filename.%modelname(%params)"
• Note: the user modified architecture should not contain any classification FC layer(e.g. the last FC layer in ArcFace). The output should be an embedding feature.

Usage

• Pretrain

   python3 train.py --method pretrain
   python3 train.py --method generator --pretrain_file %backbone_path

Note: 'pretrain' means training backbone(contraining head params) and save [backbone, head] model; 'generator' means training generator and save [backbone, head, generator] models.

• Training from scrach

  • VirClass

     python3 train.py --method pretrain
     python3 train.py --method virclass --pretrain_file %backbone_path

  • VirFace

     python3 train.py --method pretrain
     python3 train.py --method generator --pretrain_file %backbone_path
     python3 train.py --method virface --pretrain_file %generator_path

• Training from pretrained model

  • VirClass

     python3 train.py --method virclass --pretrain_file %backbone_path

  • VirFace

     python3 train.py --method virface --pretrain_file %generator_path

  Note: this pretrained file should contain [backbone, head, generator]

• Arguments setting

  method: choose training stage. [pretrain, generator, virclass, virface]

  label_batch_size: label data batch size.

  unlabel_batch_size: unlabel data batch size.

  arch: choose architecture of backbone. [resnet18, resnet34, resnet50, resnet101, resnet152, usr]

  feat_len: length of embedding feature.

  num_ids: number of identities in labeled dataset.

  gen_num: the number of generated features via generator.

  KL: weight of KL loss.

  L2: weight of MSE loss.

  resume: resume flag.

  resume_file: resume checkpoint path.

  pretrained_file: pretrained checkpoint path. This is necessary if method is not "pretrain".

  tensorboard: use tensorboard or not.

  snapshot_prefix: path to save checkpoint.

  eval: whether evaluate for each epoch on LFW, CFP-FF, CFP-FP

Citation

If our paper helps your research, please cite it in your publications:

@inproceedings{li2021virface,
  title={VirFace: Enhancing Face Recognition via Unlabeled Shallow Data},
  author={Li, Wenyu and Guo, Tianchu and Li, Pengyu and Chen, Binghui and Wang, Biao and Zuo, Wangmeng and Zhang, Lei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={14729--14738},
  year={2021}
}