CoACT

Official implementation of our paper:

Consistency-Guided Asynchronous Contrastive Tuning for Few-Shot Class-Incremental Tuning of Foundation Models
Shuvendu Roy, Elham Dolatabadi, Arash Afkanpour, and Ali Etemad
> Transactions on Machine Learning Research (TMLR 2025)

---

Overview

Getting started

Installing the dependencies

pip install -r requirements.txt

Prepare Data

14 out of 16 datasets will be downloaded automatically from huggingface datasets. For CUB-200, and miniImageNet please refer to CEC and downlaod it from here.

Extract the datasets to ./data folder.

Experiments

Resutls

Baseline: Prototype Tuning

for dataset_name in sun397 dtd voc2007 stanford_cars resisc45 oxford_pets oxford_flowers gtsrb fgvc_aircraft eurosat country211 caltech101 cifar100 cub200 food101 mini_imagenet; do

  python train.py \
    --update_base_classifier_with_prototypes True \
    --epochs_base 0 \
    --num_seeds 3 \
    --shot 10 \
    --result_key baseline \
    --dataset "${dataset_name}"

done

CoACT (Ours)

for dataset_name in sun397 dtd voc2007 stanford_cars resisc45 oxford_pets oxford_flowers gtsrb fgvc_aircraft eurosat country211 caltech101 cifar100 cub200 food101 mini_imagenet; do

  python train.py \
    --update_base_classifier_with_prototypes False \
    --start_training_with_prototypes True \
    --moco_loss_factor 1.0 \
    --epochs_base 50 \
    --num_seeds 3 \
    --num_views 2 \
    --shot 10 \
    --adapt_blocks 12 \
    --hf_model_checkpoint google/vit-base-patch16-224-in21k \
    --pet_cls LoRA \
    --result_key coact_fscit \
    --dataset "${dataset_name}" \
    --incft True

done

CoACT also shows superior performance than the prior works on conventional FSCIL setup. Following is the comparison of the results on the CIFAR-100 datasset.

Run

for dataset_name in cifar100 cub200 mini_imagenet; do

  python train.py \
    --fsl_setup "FSCIL" \
    --shot 5 \
    --epochs_base 25 \
    --incft_layers "classifier+pet" \
    --start_training_with_prototypes True \
    --pet_cls LoRA \
    --incft True \
    --result_key act_fscil \
    --num_seeds 3 \
    --dataset "${dataset_name}"

done

Acknowledgement

We would like to thank the authors of the following repositories to make their code public: MoCo, SAVC, and SupCon.

Citing CoACT

If you find our work useful to your research, please cite our paper:

@article{CoACT,
  title={Consistency-Guided Asynchronous Contrastive Tuning for Few-Shot Class-Incremental Tuning of Foundation Models},
  author={Shuvendu Roy, Elham Dolatabadi, Arash Afkanpour, Ali Etemad},
  journal={Transactions on Machine Learning Research},
  year={2025}
}