Please follow these instructions to install the environments needed with docker. After that, you can directly train and test our implementation provided in the docker environment.

This code is implemented in Unix system with CUDA 9.0.

In your conda virtual environment, install pytorch 1.1.0 following this

conda install pytorch==1.1.0 torchvision cudatoolkit=9.0 -c pytorch

After that, using pip to install all dependent packages

pip install -r requirements.txt

This is all you need for training. However, to evaluate our model by simulation, it is required to install ngsim environment. If you have some trouble with this installation, please contact us!

All setups for the parameters of training, model, dataset are in file config.py. To train our model with default parameters, simply run:

python train.py

We provide a simulation to evaluate our model compared to the baseline AGen. The simulation environment is similar to ngsim_env.

python test.py --test_mode simulation --test_datapath datasets/ngsim_22agents.h5 --ckpt_path pretrained/checkpoint-epoch100.ckpt --use_multi_agents --n_procs 5 --sim_max_obs 20

Our simulation takes around 6-7 hours to finish when only using 1 process. To speed up time, we recommend to use processes as much as possible depending on your hardware. In our cases, we used 5 processes and took around 1.5 hours to finish.

To visualize some samples of generated trajectories, follow this command

python test.py --test_mode visualization --test_datapath datasets/ngsim_22agents.h5 --use_multi_agents --ckpt_path pretrained/checkpoint-epoch100.ckpt --max_obs_length 150 --save_viz_figures

Simply run nvidia-smi while the program is running. The results might vary depending on which kind of NVIDIA card.

The CUDA support for the baseline AGen is not installed in docker environment because of package conflict.

The models part of this code is mainly based on latent-ODE. We specially thank the authors for useful code!