BT-2

Research Code for "BT^2: Backward-compatible Training with Basis Transformation" (https://arxiv.org/abs/2211.03989).

Code adapted from https://github.com/apple/ml-fct.

Requirements

We suggest using Conda virtual environments, please run:

conda env create -f environment.yml
conda activate sm86

Dataset Preparation

Make dataset and checkpoint directories.

mkdir data_store
mkdir checkpoints

Cifar 100

Please refer to https://www.cs.toronto.edu/~kriz/cifar.html for downloading Cifar 100.

Imagenet 1k

Please refer to https://www.image-net.org/challenges/LSVRC/2012/index.php for downloading the Imagenet 1k.

Example Experiments on Cifar 100

We provide training and evaluation experiment configurations for Cifar 100 in ./configs. The following commands are backward compatible training experiments from ResNet50 to ResNet50 (with data change of 50 classes to 100 classes).

Train Old Backbone Model

python train_backbone.py --config configs/cifar100_backbone_old.yaml

Train Independent New Backbone Model

python train_backbone.py --config configs/cifar100_backbone_new.yaml

Train Backward-compatible New Model with Basis Transformation

python train_feature_transfer.py --config configs/cifar100_transfer.yaml

Train Backward-compatible New Model with BCT (https://arxiv.org/abs/2003.11942)

python train_BCT.py --config configs/cifar100_BCT.yaml

Evaluation of Old/Old (query feature/ gallery feature)

python eval.py --config configs/cifar100_eval_old_old.yaml

Evaluation of New/New (query feature/ gallery feature)

python eval.py --config configs/cifar100_eval_new_new.yaml

Evaluation of New/Old (query feature/ gallery feature)

python eval.py --config configs/cifar100_eval_old_new.yaml

Example Experiments on Imagenet

We provide training and evaluation experiment configurations for Imagenet 1k in ./configs. Commands are similar to commands used for experiments in Cifar 100.

Checkpoints and results:

We provide trained checkpoints using config files in example configurations and in the paper here.

Example experiment results on Cifar100.

Method	Setting	TOP1	TOP5	meanAP
Independent	$\phi_{old}/\phi_{old}$ $\phi_{new}/\phi_{old}$ $\phi_{new}/\phi_{new}$	33.6 0.8 62.7	55.4 4.9 74.6	24.4 1.5 49.9
BCT	$\phi_{new}/\phi_{old}$ $\phi_{new}/\phi_{new}$	25.0 60.0	62.1 71.9	24.7 47.3
$BT^2$(ours)	$\phi_{new}/\phi_{old}$ $\phi_{new}/\phi_{new}$	38.7 62.4	64.6 75.1	27.7 50.5

Example experiment results on Imagenet1k.

Method	Setting	TOP1	TOP5	meanAP
Independent	$\phi_{old}/\phi_{old}$ $\phi_{new}/\phi_{old}$ $\phi_{new}/\phi_{new}$	40.9 0.1 67.9	55.8 0.5 81.4	33.6 0.2 52.3
BCT	$\phi_{new}/\phi_{old}$ $\phi_{new}/\phi_{new}$	44.3 65.3	66.4 80.0	34.6 54.0
$BT^2$(ours)	$\phi_{new}/\phi_{old}$ $\phi_{new}/\phi_{new}$	44.4 66.6	65.7 81.1	35.0 54.6

* Some results are different from the paper due to sensitivity to hyperparameters and random seeds.

Experiment results of sequence updates on Imagenet1k.

Method	Setting	TOP1	TOP5	meanAP
Independent	$\phi_{alex}/\phi_{alex}$ $\phi_{vgg}/\phi_{vgg}$ $\phi_{res}/\phi_{res}$ $\phi_{vit}/\phi_{vit}$	46.6 63.2 67.9 78.0	66.3 79.0 81.4 87.5	29.1 49.6 52.3 72.4
BCT	$\phi_{vgg}/\phi_{alex}$ $\phi_{vgg}/\phi_{vgg}$ $\phi_{res}/\phi_{alex}$ $\phi_{res}/\phi_{vgg}$ $\phi_{res}/\phi_{res}$ $\phi_{vit}/\phi_{alex}$ $\phi_{vit}/\phi_{vgg}$ $\phi_{vit}/\phi_{res}$ $\phi_{vit}/\phi_{vit}$	54.4 58.4 46.0 48.9 64.3 54.9 57.5 70.3 73.9	74.1 75.4 71.9 75.2 79.1 82.0 84.1 85.1 86.0	36.2 47.0 30.6 44.4 52.7 36.3 50.5 57.0 65.8
$BT^2$(ours)	$\phi_{vgg}/\phi_{alex}$ $\phi_{vgg}/\phi_{vgg}$ $\phi_{res}/\phi_{alex}$ $\phi_{res}/\phi_{vgg}$ $\phi_{res}/\phi_{res}$ $\phi_{vit}/\phi_{alex}$ $\phi_{vit}/\phi_{vgg}$ $\phi_{vit}/\phi_{res}$ $\phi_{vit}/\phi_{vit}$	56.5 61.0 56.7 61.5 66.6 57.9 62.5 72.0 75.6	75.6 77.2 78.5 80.8 80.8 83.5 86.5 87.0 87.4	37.1 47.5 37.2 50.6 56.8 37.6 52.7 60.6 68.0

Cite our paper

@misc{zhou2023bt2,
      title={$BT^2$: Backward-compatible Training with Basis Transformation}, 
      author={Yifei Zhou and Zilu Li and Abhinav Shrivastava and Hengshuang Zhao and Antonio Torralba and Taipeng Tian and Ser-Nam Lim},
      year={2023},
      eprint={2211.03989},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}