Fine-grained Classes and How to Find Them

Matej Grcić, Artyom Gadetsky, Maria Brbić

Project page| arXiv | BibTeX

This repo contains PyTorch implementation of the FALCON method. FALCON is a method that discovers fine-grained classes in coarsely labeled data without any supervision at the fine-grained level. For more details please check our paper Fine-grained Classes and How to Find Them (ICML '24).

Installation

Setup the conda environment and install the required packages by running the following commands:

conda env create -f environment.yml
conda activate falcon
./scripts/setup.sh

Data preparation

We use the standard ImageNet-1k and CIFAR100 datasets. Download the datasets and place them in the desired directory.

For single-cell data, we provide a download and preprocess script:

./scripts/download_sc_data.sh /path/to/datasets_dir

Export the dataset paths:

export CIFAR100_ROOT=/PATH/TO/CIFAR100
export IMAGENET_ROOT=/PATH/TO/IMAGENET
export PBMC_ROOT=/PATH/TO/PREPROCESSED/PBMC

Training

To train FALCON, simply run one of the available scripts. For example, to train FALCON on the Living17 dataset, run:

./script/breeds/living17.sh /path/to/output_dir

Notes:

• You might need to make the script executable by running chmod +x ./script/breeds/living17.sh
• The training can be resumed by proving the --resume flag to main.py.

Evaluation

To evaluate the model, run the following command with the desired config, model path, and output directory:

python main.py \
  --cfg_file configs/breeds/coarse2fine/living17.yaml \
  --override_cfg OUTPUT_DIR /path_to_eval_dir \
  --eval_only --model /path_to_model.pth

Checkpoints

FALCON checkpoints can be found here.

Citation

If you find our code useful, please consider citing:

@inproceedings{
grcic24icml,
title={Fine-grained Classes and How to Find Them},
author={Grci’c, Matej and Gadetsky, Artyom and Brbi’c, Maria},
booktitle={International Conference on Machine Learning},
year={2024},
}