This repository contains the implementation for the paper "H-SPLID: HSIC-based Saliency Preserving Latent Information Decomposition" (NeurIPS 2025).

You can install the project dependencies using either Conda or pip:

git clone https://github.com/neu-spiral/H-SPLID.git
cd h-splid

# Pip:
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Conda:
conda env create -f environment.yml
conda activate hsplid

The main training script can be run with:

python train.py --config-name=<config_name>

Configuration is managed through Hydra, with config files located in the configs/ directory.

To reproduce the results of our COCO experiments, follow these steps:

1. Adapt the configurations to your setup:

   • Set your Weights & Biases entity (or disable W&B logging) in configs/user/user.yaml.
   • Specify the path for downloading the COCO dataset in configs/dataset/coco-animals.yaml.

2. Pretrained vanilla model:

   • Option A: Use the provided pretrained model from assets/models/coco-animals-standard-resnet18.pt.

   • Option B: Train your own vanilla model by running:

     python train.py --config-name=coco_pretrain

     If you train your own model, update the pretrained_model_path in configs/coco_hsplid.yaml to point to your trained model.

3. Select and run your experiment:

   • The experiment can be selected by setting the experiment flag in the defaults section of configs/coco_hsplid.yaml to one of the experiment configs in configs/experiment/coco-animals/resnet18/.
   • Run the experiment:

   python train.py --config-name=coco_hsplid

4. Run the attacks using configs/coco_attacks.yaml:

   • Update W&B settings in the config file.
   • In the Models section, list each model's description and path.
   • The script will loop over the specified seeds and compute the averaged results:

   python attacks.py --config-name=coco_attacks

To reproduce the results of our ImageNet experiments, follow these steps:

1. Download datasets and benchmarks:

   • Download the ImageNet-1K dataset and the desired saliency benchmarks (links provided in the paper).

2. Adapt the configurations to your setup:

   • Specify the path to the ImageNet-1K dataset in configs/dataset/imagenet.yaml.
   • Specify the path to the ImageNet9 dataset in configs/dataset/imagenet9.yaml. You can also select the split you want to use for benchmarking with transfer.py.
   • Specify the path to the CounterAnimal dataset in configs/dataset/counteranimal.yaml. You can also select the split you want to use for benchmarking with transfer.py.
   • Set your Weights & Biases entity (or disable W&B logging) in configs/user/user.yaml.

3. Train models

   • The experiment can be selected by setting the experiment flag in the defaults section of configs/imagenet_hsplid.yaml to one of the experiment configs in configs/experiment/imagenet/resnet50/.
   • Run the experiment:

   python train.py --config-name=imagenet_hsplid

4. Evaluate models on benchmark datasets:

   • Set model_path in configs/imagenet_transfer.yaml to the desired model path.
   • Set the dataset flag in the defaults section to one of the desired datasets: imagenet9, counteranimal and update the split.
   • Run:

   python transfer.py --config-name=imagenet_transfer

To reproduce the results of our ISIC2017 experiments, follow these steps:

1. Adapt the configurations to your setup:

   • Set your Weights & Biases entity (or disable W&B logging) in configs/user/user.yaml.
   • Specify the path for the ISIC2017 dataset in configs/dataset/isic2017.yaml.

2. Pretrained vanilla model:

   • Option A: Use the provided pretrained model from assets/models/isic2017-standard-resnet50.py.

   • Option B: Train your own vanilla model by running:

     python train.py --config-name=isic_pretrain

   If you train your own model, update the pretrained_model_path in configs/isic_hsplid.yaml to point to your trained model.

3. Select and run your experiment:

   • The experiment can be selected by setting the experiment flag in the defaults section of configs/isic_hsplid.yaml to one of the experiment configs in configs/experiment/isic2017/resnet50/.
   • Run the experiment:

   python train.py --config-name=isic_hsplid

4. Run the attacks using configs/isic_attacks.yaml:

   • Update W&B settings in the config file.
   • In the Models section, list each model's description and path.
   • The script will loop over the specified seeds and compute the averaged results:

   python attacks.py --config-name=isic_attacks

If you use this code in your research, please cite (NeurIPS citation coming soon):

@misc{miklautz2025hsplidhsicbasedsaliencypreserving,
      title={H-SPLID: HSIC-based Saliency Preserving Latent Information Decomposition}, 
      author={Lukas Miklautz and Chengzhi Shi and Andrii Shkabrii and Theodoros Thirimachos Davarakis and Prudence Lam and Claudia Plant and Jennifer Dy and Stratis Ioannidis},
      year={2025},
      eprint={2510.20627},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2510.20627}, 
}