conda create --name ssat python=3.7
conda activate ssat
conda install pytorch==1.5.1 torchvision cudatoolkit=10.2 -c pytorch #
# install argoverse api
pip install git+https://github.com/argoai/argoverse-api.git
#or from their website: https://github.com/argoverse/argoverse-api#installation
# install others dependancy
pip install scikit-image IPython tqdm ipdb
Download the dataset Argoverse 1 and store them under the folder ./dataset
python train_adv_ssat.py -m _ssat_model --resume=basenet.ckpt --att_pattern=ade
For the sake of simplicity, the uploaded version generates the adversarial trajectories during runtime, which is quite slow. In fact, you could add several lines to store the generated adversarial trajectories and load them during adversarial training.