Explainable Trajectory Prediction

This repository contains the code for explainable trajectory prediction based on Shapley values.

Data Preparation

The processing code/files of the ETH-UCY, nuScenes and SDD can be found at:

ETH-UCY, SDD for Trajectron++: https://github.com/StanfordASL/Trajectron-plus-plus/blob/master/experiments/pedestrians/process_data.py
nuScenes for Trajectron++: https://github.com/StanfordASL/Trajectron-plus-plus/blob/master/experiments/nuScenes/process_data.py
SDD for PECNet: https://github.com/HarshayuGirase/PECNet/tree/master/social_pool_data

To process the SportVU dataset for the two frameworks:

python tools/prepare_data.py
python tools/process_trajectron_data.py
python tools/process_pecnet_data.py

The raw data can be accessed at: https://github.com/linouk23/NBA-Player-Movements/tree/master/data/2016.NBA.Raw.SportVU.Game.Logs

Thirdparty

Download the following frameworks and place them under:

thirdparty/Trajectron-plus-plus: https://github.com/StanfordASL/Trajectron-plus-plus
thidparty/PECNet: https://github.com/HarshayuGirase/PECNet

Pre-trained Models

All pre-trained models are provided under models/.

Models Testing

Examples to test models:

# Testing Traj++Edge on ETH
env PYTHONPATH=src python bin/test_trajectron.py models/trajectron/eth-ucy/eth_edge 100 data/eth-ucy/eth_test.pkl cpu 
# Testing PECNet on SDD
env PYTHONPATH=src python bin/test_pecnet.py thirdparty/PECNet/saved_models/PECNET_social_model1.pt thirdparty/PECNet/social_pool_data/test_all_4096_0_100.pickle

To test the models without interaction, append --without_neighbours to the call.

Model Training

An example call to train Trajectron++Edge:

# Training Traj++Edge on SportVU 
env PYTHONPATH=src python bin/train_trajectron.py --config models/trajectron/sport/edge/config.json --data_dir data/sport --train_data_dict trajectron_train.pkl --log_dir logs --log_tag sport_traj++Edge --train_epochs 20 --save_every 5 --deeper_action --late_fusion

Shapley Values Estimation

Example calls to compute Shapley values:

# Shapley values of Traj++Edge on the first scene of SportVU
env PYTHONPATH=src python bin/compute_shapley_values_trajectron.py models/trajectron/sport/edge 20 data/sport/trajectron_test.pkl cuda:0 home nll zero 0 results/sport/Traj++Edge --random_node_types home guest

# Shapley values of PECNet on the first scene of SportVU
env PYTHONPATH=src python bin/compute_shapley_values_pecnet.py models/pecnet/sport.pt data/sport/pecnet_test.pkl 0 results/sport/pecnet

The above calls estimate the Shapley values per scene, to merge the results over all scenes of a dataset:

python tools/merge_results.py results/sport Traj++Edge nll results

Plotting

An example call to plot the aggregated Shapley values by comparing two models:

python tools/plot_shapley_values.py --names Traj++ Traj++Edge --paths results/Traj++_nll.pkl results/Traj++Edge_nll.pkl --output_path results/SportVU_nll.png

To plot the local analysis (per scenario) of the SportVU:

python tools/plot_scenarios.py results/sport/Traj++Edge_home_0_nll.pkl results/court.png

where the second argument point to the image of the scene and can be accessed at: https://github.com/linouk23/NBA-Player-Movements/

Unit Tests

The following script checks a set of unit tests.

./local_run_tests.sh

Security

See CONTRIBUTING for more information.

License

This project is licensed under the Apache-2.0 License.

leesunfreshing / explainable-trajectory-prediction Goto Github PK