Model Developmental Safety: A Safety-Centric Method and Applications in Vision-Language Models

This repo is to provide implementation of the algorithm proposed in paper Model Developmental Safety: A Safety-Centric Method and Applications in Vision-Language Models and reproduce the experimental results.

1. Setup Environment

• Git clone this repo

git clone https://github.com/GangLii/DevSafety

• Create conda environment

conda create -n DevSafety python=3.10
conda activate DevSafety
cd ./DevSafety
pip install -r requirements.txt

2. Datasets and Base Models

1. Download BDD100K dataset from the website and Places365 dataset from here
2. Download base models for BDD100K experiments from here. These models are pretrained on BDD100K dataset.
3. Download retrieved Laion data from here
4. Download csvs file for datasets splits from here.
5. Unzip all downloaded files

3. Link Downloaded Files

Link downloaded files or move downloaded files to corresponding folders. To generate system links to downloaded files, go to the cloned repo folder, and run below commands.

ln -s path_to_folder_for_bdd100k/bdd100k               ./data/datasets/bdd100k
ln -s path_to_folder_for_places365/places365           ./data/datasets/places365
ln -s path_to_folder_for_laion_foggy/laion_foggy       ./data/datasets/laion_foggy
ln -s path_to_folder_for_laion_overcast/laion_overcast ./data/datasets/laion_overcast
ln -s path_to_folder_for_laion_tunnel/laion_tunnel     ./data/datasets/laion_tunnel
ln -s path_to_folder_for_laion_dressing_room/laion_dressing_room   ./data/datasets/laion_dressing_room
ln -s path_to_folder_for_laion_negative/laion_negative   ./data/datasets/laion_negative
ln -s path_to_folder_for_base_models/base_models       ./data/base_models
ln -s path_to_folder_for_csvs/csvs                     ./data/csvs

4. Run the experiments

Now, we are ready to run experiments. For example,

bash train_foggy.sh.

We used 2x40G GPUs for BDD100k experiments, which costs around 12 hours for each run. We used 4x48G GPUs for Places365 experiments, which costs around 44 hours for each run.

Credits: The pipline of our code is based on https://github.com/locuslab/FLYP. We thank the authors for making their code publicly available.

Contact

If you have any questions, please open a new issue or contact Gang Li via [email protected]. If you find this repo helpful, please cite the following paper:

@article{li2024model,
  title={Model Developmental Safety: A Safety-Centric Method and Applications in Vision-Language Models},
  author={Li, Gang and Yu, Wendi and Yao, Yao and Tong, Wei and Liang, Yingbin and Lin, Qihang and Yang, Tianbao},
  journal={arXiv preprint arXiv:2410.03955},
  year={2024}
}