OCA

Code for "Backward-Compatible Aligned Representations via an Orthogonal Transformation Layer" (https://arxiv.org/abs/2408.08793).

Code adapted from https://github.com/apple/ml-fct and https://github.com/YifeiZhou02/BT-2

Requirements

To create the Conda virtual environment, please run:

conda env create -f environment.yml
conda activate OCA

Dataset Preparation

Make dataset, checkpoint and logs directories.

mkdir data_store
mkdir checkpoints
mkdir logs_result

Cifar100

Please refer to https://www.cs.toronto.edu/~kriz/cifar.html for downloading Cifar100.

Imagenet 1k

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

Example Experiments on Cifar100

We provide training and evaluation experiment configurations for Cifar100 in ./configs. The following commands are training experiments from ResNet50 to ResNet50 (with data change of 50 classes to 100 classes).

Train and Eval script

bash run_train_eval.sh

The results are saved in logs_result folder.

Cite our paper

@misc{ricci2024backwardcompatiblealignedrepresentationsorthogonal,
      title={Backward-Compatible Aligned Representations via an Orthogonal Transformation Layer}, 
      author={Simone Ricci and Niccolò Biondi and Federico Pernici and Alberto Del Bimbo},
      year={2024},
      eprint={2408.08793},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.08793}, 
}