Skip to content

Home

Jump to bottom
yuleung edited this page Aug 23, 2023 · 1 revision

Welcome to the MixBCT wiki!

MixBCT: Towards Self-Adapting Backward-Compatible Training

Introduction

Implementation of MixBCT: Towards Self-Adapting Backward-Compatible Training(Ours) , L2 and other SOTA methods: UniBCT, NCCL, BCT, AdvBCT

L2: Simple L2 constraint between old feature and new feature

BCT: Towards Backward-Compatible Representation Learning (CVPR2020)

UniBCT: Towards Universal Backward-Compatible Representation Learning (AAAI2022)

NCCL: Neighborhood Consensus Contrastive Learning for Backward-Compatible Representation (IJCAI2022)

AdvBCT: Boundary-Aware Backward-Compatible Representation via Adversarial Learning in Image Retrieval (CVPR2023)

Dataset

  • Training dataset: MS1M-V3 (ms1m-retinaface) 5179510 images with 93431 IDs
  • Eval dataset: IJB-C

The download link of the datasets can be find in [https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_]

Code Structure

  • The main-dir(./) is used for train the Old model
  • ./BCT_methods/ --- The methods which summarized MixBCT, UniBCT, NCCL, BCT, AdvBCT and L2.
  • ./tools/ --- dataset split code, some preprocessing operations code and IJB-C evaluation code

This code based on the project insightface, We maintain separate directories for each method to enhance clarity and facilitate reproducibility.

Note: We fixed the random seed in the main file for training, and this will significantly reduce the speed of training. You can speed up the training by comment out following two lines in the main file:

#torch.backends.cudnn.benchmark = False;
#torch.backends.cudnn.deterministic = True

However, it will result in slight randomness of the results.

Training Flow ---- An Example:

Step-1

Train the old model use the arcface loss.

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1  --master_port=22222 train_old_arc.py configs/f512_r18_arc_class30.py

Or train the old model use the softmax loss.

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1  --master_port=22222 train_old_softmax.py configs/f128_r18_softmax_class30.py

Step-2 ----（preprocessing operations）used in MixBCT、NCCL

Get the feature of the dataset consist of 'class70' images.

python tools/get_feature/get_avg_feature.py configs/f128_r18_softmax_class30.py --SD f128_r18_softmax_class70

Step-2 ----（preprocessing operations）used in BCT、UniBCT

Get the avg feature of the dataset consist of 'class70' images(based on Step-2).

python tools/get_feature/get_avg_feature.py  --SD f128_r18_softmax_class70

Step-2 ----（preprocessing operations）used in MixBCT

Get the denoised feature of the dataset consist of 'class70' images(based on Step-2).

python tools/get_feature/denoise_credible.py --T 0.9 --SD f128_r18_softmax_class70

Step-3

Train the New-Model by MixBCT

cd BCT_Methods/MixBCT/
python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1  --master_port=22222 train.py configs/OPclass_ms1mv3_r18_to_r50_MixBCT_softmax_to_arc_f128.py

Or train the New-Model by NCCL

cd BCT_Methods/NCCL/
python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1  --master_port=22222 train.py configs/OPclass_ms1mv3_r18_to_r50_NCCL_softmax_to_arc_f128.py

Or train the New model by other methods

cd BCT_Methods/#Other Methods/
python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1  --master_port=22222 train.py configs/OPclass_ms1mv3_r18_to_r50_(Othermethods)_softmax_to_arc_f128.py

Step-4

IJB-C evaluation

**self-test 1:1**
python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) 
**self-test 1:N**
python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) -N 
**cross-test 1:1**
python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) -m_old=#The path of 'Old_model.pt' -old_net=#The backbone of Old model(r18,r50,vit...) 
**cross-test 1:N**
python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) -m_old=#The path of 'Old_model.pt' -old_net=#The backbone of Old model(r18,r50,vit...) 

# IJB-C evaluation-Using nohup
#CUDA_VISIBLE_DEVICES=0 nohup python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) >>#Result save path 2>&1 &
#CUDA_VISIBLE_DEVICES=1 nohup python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) -N >>#Result save path 2>&1 &
#CUDA_VISIBLE_DEVICES=2 nohup python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) -m_old=#The path of 'Old_model.pt' -old_net=#The backbone of Old model(r18,r50,vit...) >>#Result save path 2>&1 &
#CUDA_VISIBLE_DEVICES=3 nohup python tools/ijbc_eval/ijbc_eval.py -m=#The path of 'New_model.pt' -net=#The backbone of Nld model(r18,r50,vit...) -m_old=#The path of 'Old_model.pt' -old_net=#The backbone of Old model(r18,r50,vit...) >>#Result save path 2>&1 &

Citation

If this repository helps your research, please cite our paper:

@article{liang2023mixbct,
  title={MixBCT: Towards Self-Adapting Backward-Compatible Training},
  author={Liang, Yu and Zhang, Shiliang and Wang, Yaowei and Xiao, Sheng and Li, Kenli and Wang, Xiaoyu},
  journal={arXiv preprint arXiv:2308.06948},
  year={2023}
}
Clone this wiki locally