OpenBCT is a Non Official Reimplementation for Towards Backward-Compatible Representation Learning (CVPR 2020 oral), which aims at achieving compatibility across different models.
- Baselines for Backward Compatible Training
- Influence Loss with old classifier
- Pseudo old classifier generation
- Python >= 3.6
- Pytorch >= 1.2.0
- torchvision == 0.2.1
- numpy
- sklearn
- easydict
This code conducts training and evaluation on ImageNet LSVRC 2012 and Places365(easy directory structure) datasets. Please be noticed that, due to the privacy issue, this code will NOT provide the train and test code for face datasets in our CVPR 2020 paper.
After you download these two datasets, extract them and make sure there are ./train and ./val folders inside. You can find the training image lists we use from Google Drive Link or Weiyun Link.
For training an old model without any regularization,
python main.py your_dataset_dir --train-img-list imgnet_train_img_list_for_old.txt -a resnet18 For training a new model with infulence loss (old classifier regularization),
python main.py your_dataset_dir --train-img-list imgnet_train_img_list_for_new.txt -a resnet50 --old-fc your_old_fc_weights_dir --n2o-map ./imgnet_new2old_map.npyFor training a new model with L2 regression loss (one of the compared baseline),
python main.py your_dataset_dir --train-img-list imgnet_train_img_list_for_new.txt -a resnet50 --old-arch resnet18 --old-checkpoint your_old_model_dir --l2 --use-featFor cross test between two models,
python main.py your_dataset_dir -a resnet50 --pretrained --checkpoint your_new_model_dir --old-fc your_old_fc_weights_dir --use-feat -e --cross-eval --old-arch resnet18 --old-checkpoint your_old_model_dir For self test with single model,
python main.py your_dataset_dir -a resnet50 --pretrained --checkpoint your_model_dir --use-feat -eFor training an old model without any regularization,
python main.py your_dataset_dir --dist-url 'tcp://127.0.0.1:FREEPORT' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --train-img-list imgnet_train_img_list_for_old.txt -a resnet18 For training a new model with infulence loss,
python main.py your_dataset_dir --dist-url 'tcp://127.0.0.1:FREEPORT' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --train-img-list imgnet_train_img_list_for_new.txt -a resnet50 --old-fc your_old_fc_weights_dir --n2o-map ./imgnet_new2old_map.npyFor training a new model with L2 regression loss (one of the compared baseline),
python main.py your_dataset_dir --dist-url 'tcp://127.0.0.1:FREEPORT' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --train-img-list imgnet_train_img_list_for_new.txt -a resnet50 --old-arch resnet18 --old-checkpoint your_old_model_dir --l2 --use-featNote: This is for single machine, multi GPUs
Influence loss results on ImagNet are listed as below. Please be noted that, in testing, we use feature L2 distance for Top-k accuracy computing. More results are on the way.
| Old Model | New Model | Val Set | Top-1 Acc | Top-5 Acc |
|---|---|---|---|---|
| resnet18 (50%) | resnet18 (50%) | ImageNet Val | 39.5% | 60.0% |
| resnet18 (50%) | resnet50-BCT (100%) | ImageNet Val | 42.2% | 65.5% |
| resnet18 (50%) | resnet50-BCT-normed-clsfier (100%) | ImageNet Val | 46.8% | 66.7% |
| resnet18 (50%) | resnet50-L2 (100%) | ImageNet Val | 13.0% | 32.8% |
| resnet18 (50%) | resnet50-Triplet (100%) | ImageNet Val | 42.9% | 63.3% |
| resnet18 (50%) | resnet50-Contra (100%) | ImageNet Val | 42.7% | 63.2% |
| resnet50-L2 (100%) | resnet50-L2 (100%) | ImageNet Val | 43.8% | 64.4% |
| resnet50-Triplet (100%) | resnet50-Triplet (100%) | ImageNet Val | 53.7% | 74.3% |
| resnet50-Contra (100%) | resnet50-Contra (100%) | ImageNet Val | 57.0% | 76.3% |
| resnet50-BCT (100%) | resnet50-BCT (100%) | ImageNet Val | 55.6% | 76.6% |
| resnet50 (100%) | resnet50 (100%) | ImageNet Val | 66.3% | 84.0% |
| Old Model | New Model | Val Set | Top-1 Acc | Top-5 Acc |
|---|---|---|---|---|
| resnet18 (50%) | resnet18 (50%) | Places365 Val | 27.0% | 55.9% |
| resnet18 (50%) | resnet50-BCT (100%) | Places365 Val | 27.5% | 57.8% |
| resnet50-BCT (100%) | resnet50-BCT (100%) | Places365 Val | 32.9% | 62.2% |
| resnet50 (100%) | resnet50 (100%) | Places365 Val | 35.1% | 64.0% |
In this table, x% denotes the training data usage amount.
The code is based on Open-ReID and Pytorch-ImageNet-Example. Thank these researchers for sharing their great code!
If this code helps your research or project, please cite
@inproceedings{shen2020towards,
title={Towards backward-compatible representation learning},
author={Shen, Yantao and Xiong, Yuanjun and Xia, Wei and Soatto, Stefano},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={6368--6377},
year={2020}
}
If you have any question, please feel free to contact
Yantao Shen: ytshen@link.cuhk.edu.hk