Mixup in Agent Forecasting

This codebase is the implementation for the final project in Fall 2022 CS 182 Final Project. Our paper is Mixup in Agent Forecasting.

The majority of the codebase is from Trajectron++. Its paper link is Trajectron++: Dynamically-Feasible Trajectory Forecasting With Heterogeneous Data

by Tim Salzmann*, Boris Ivanovic*, Punarjay Chakravarty, and Marco Pavone (* denotes equal contribution).

Cloning

When cloning this repository, make sure you clone the submodules as well, with the following command:

git clone --recurse-submodules <repository cloning URL>

Alternatively, you can clone the repository as normal and then load submodules later with:

git submodule init # Initializing our local configuration file
git submodule update # Fetching all of the data from the submodules at the specified commits

Environment Setup

First, we'll create a conda environment to hold the dependencies.

conda create --name trajectron++ python=3.6 -y
source activate trajectron++
pip install -r requirements.txt

Data Setup

We've already included preprocessed data splits for the ETH and UCY Pedestrian datasets in this repository, you can see them in experiments/pedestrians/raw. In order to process them into a data format that our model can work with, execute the follwing.

cd experiments/pedestrians
python process_data.py # This will take around 10-15 minutes, depending on your computer.

Replicate Experiments

We have already written all the bash code you need to run in order to obtain the baseline and two different kinds of mixup models. You need first go into trajectron directory. To train all the baseline results, run

bash run_baseline.sh

To train all the mixup results on train set, run

bash run_mixup.sh

You can find all trained model in experiments/pdestrians. There are two types of Trajectron++ baseline models. One is with dynamic integration, whose model suffix is attention_radius_3, and the other is without, whose suffix is vel.

To evaluate all the trained models (both baseline ones and mixup ones) on test set, run

bash evaluate_predictor.sh

To analyze the results from all the models, run

bash run_analysis.sh

It will then display the folder name of that generic model along with the result in a table. You can also retrieve that result in the file called {folder_name}/best_of_20.csv.

If you want understand the hyperparameters in the training bash code, I strongly recommend take a look at Trajectron++ codebase