We propose a novel method for generating noise-free explanations for ViT and evaluate our method on driving action prediction task.

Our framework consists of two stages: self-supervised training and supervised multi-label classification fine-tuning.

conda (Recommended) - Clone the repository and then create and activate a the conda environment using the provided environment definition:

conda env create -f SNNA.yaml
conda activate SNNA

The data used in this project is the BDD100k and BDD-OIA The data should be downloaded and placed in the dataset directory. The directory structure should look like this:

|-- dataset
    |-- BDD100k
        |-- images
            |-- 100k
                |-- train
                |-- val
                |-- test
    |-- BDD-OIA
        |-- test
        |-- train
        |-- val
        test.json
        train.json
        val.json

To train the model, run the following command:

1. Self-supervised Fine-tuning the model on BDD100k

python -m torch.distributed.launch --nproc_per_node=1 main_dino.py --patch_size 8 --batch_size_per_gpu 16 --epochs 200 --saveckp_freq 10 --data_path /dataset/BDD100k/images/100k --output_dir /ckp/

2. Supervised Multi-label classification Fine-tuning the classifier on BDD-OIA

python multi_label_train.py --num_labels 4 --patch_size 8 --batch_size_per_gpu 4  --epochs 100 --pretrained_weights /ckp/backbone_200.pth --data_path /dataset/BDD-OIA/ --output_dir /ckp/

This repository is based on the following works:

• DINO
• Transformer-Explainability we thank the authors for their excellent works.

If you make use of our work, please cite our paper: