Generative Multi-Agent Behavioral Cloning

Code is written with PyTorch (Python 3.6.1).

Dataset is available from STATS. A pre-processed version is available here. Download the data into the bball_data/data/ folder.

Example Run

You can edit train_model.sh and run that script, or run train.py directly with command-line parameters.

$ ./train_model.py
$ python sample.py -t 105 -b 10 -n 10
$ python plot.py -t 105 --animate

Trained models for RNN_GAUSS (101), VRNN_SINGLE (102), VRNN_INDEP (103), and our model MACRO_VRNN (104) are included.

Files

model.py contains the models. MACRO_VRNN is our model with macro-goals.

train.py contains the training process, and can be called with train_model.sh.

sample.py is used to sample rollouts from a trained model.

plot.py is used to plot the samples as well as animate them (with --animate flag). --showmacro will display macro-goals, where applicable.

model_utils.py contains functions for sampling and calculating various losses.

bball_data\__init__.py contains the Dataset object.

bball_data\cfg.py contains constants for the data.

bball_data\macro_goals.py is the script used to extract macro-goals. Don't need to run again.

bball_data\utils.py contains the functions for plotting and animating.