Trajectron++: Dynamically-Feasible Trajectory Forecasting With Heterogeneous Data

This repository contains the code for Trajectron++: Dynamically-Feasible Trajectory Forecasting With Heterogeneous Data by Tim Salzmann*, Boris Ivanovic*, Punarjay Chakravarty, and Marco Pavone (* denotes equal contribution).

Installation

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

Then, since this project uses IPython notebooks, we'll install this conda environment as a kernel.

python -m ipykernel install --user --name trajectronpp --display-name "Python 3.6 (Trajectron++)"

Data Setup

Pedestrian Datasets

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.

nuScenes Dataset

Download the nuScenes dataset (this requires signing up on their website). Note that the full dataset is very large, so if you only wish to test out the codebase and model then you can just download the nuScenes "mini" dataset which only requires around 4 GB of space. Extract the downloaded zip file's contents and place them in the experiments/nuScenes directory. Then, download the map expansion pack (v1.1) and copy the contents of the extracted maps folder into the experiments/nuScenes/v1.0-mini/maps folder. Finally, process them into a data format that our model can work with.

cd experiments/nuScenes

# For the mini nuScenes dataset, use the following
python process_data.py --data=./v1.0-mini --version="v1.0-mini" --output_path=../processed

# For the full nuScenes dataset, use the following
python process_data.py --data=./v1.0 --version="v1.0-trainval" --output_path=../processed

In case you also want a validation set generated (by default this will just produce the training and test sets), replace line 406 in process_data.py with:

    val_scene_names = val_scenes

Model Training

Pedestrian Dataset

To train a model on the ETH and UCY Pedestrian datasets, you can execute a version of the following command from within the trajectron/ directory.

python train.py --eval_every 10 --vis_every 1 --train_data_dict <dataset>_train.pkl --eval_data_dict <dataset>_val.pkl --offline_scene_graph yes --preprocess_workers 5 --log_dir ../experiments/pedestrians/models --log_tag <desired tag> --train_epochs 100 --augment --conf <desired model configuration>

For example, a fully-fleshed out version of this command to train a model without dynamics integration for evaluation on the ETH - University scene would look like:

python train.py --eval_every 10 --vis_every 1 --train_data_dict eth_train.pkl --eval_data_dict eth_val.pkl --offline_scene_graph yes --preprocess_workers 5 --log_dir ../experiments/pedestrians/models --log_tag _eth_vel_ar3 --train_epochs 100 --augment --conf ../experiments/pedestrians/models/eth_vel/config.json

What this means is to train a new Trajectron++ model which will be evaluated every 10 epochs, have a few outputs visualized in Tensorboard every 1 epoch, use the eth_train.pkl file as the source of training data (which actually contains the four other datasets, since we train using a leave-one-out scheme), and evaluate the partially-trained models on the data within eth_val.pkl. Further options specify that we want to perform a bit of preprocessing to make training as fast as possible (--offline_scene_graph yes), use 5 threads to parallelize data loading, save trained models and Tensorboard logs to ../experiments/pedestrians/models, mark the created log directory with an additional _eth_vel_ar3 at the end, run training for 100 epochs, augment the dataset with rotations (--augment), and use the same model configuration as in the model we previously trained for the ETH dataset without any dynamics integration (--conf ../experiments/pedestrians/models/eth_vel/config.json).

If you wanted to train a model with dynamics integration for the ETH - University scene, then you would instead run:

python train.py --eval_every 10 --vis_every 1 --train_data_dict eth_train.pkl --eval_data_dict eth_val.pkl --offline_scene_graph yes --preprocess_workers 5 --log_dir ../experiments/pedestrians/models --log_tag _eth_ar3 --train_epochs 100 --augment --conf ../experiments/pedestrians/models/eth_attention_radius_3/config.json

where the only difference is the sourced model configuration (now from ../experiments/pedestrians/models/eth_attention_radius_3/config.json). Our codebase is set up such that hyperparameters are saved in a json file every time a model is trained, so that you don't have to remember what settings you use when you end up training many models in parallel!

Commands like these would be used for all of the scenes in the ETH and UCY datasets (the options being eth, hotel, univ, zara1, and zara2). The only change would be what train_data_dict, eval_data_dict, log_tag, and configuration file (conf) you wish to use.

nuScenes Dataset

To train a model on the nuScenes dataset, you can execute one of the following commands from within the trajectron/ directory, depending on the model version you desire.

In case you also want to produce the version of our model that was trained without the ego-vehicle (first row of Table 4 (b) in the paper), then run the command from the third row of the table above, but change line 132 of train.py to:

                                       return_robot=False)

CPU Training

By default, our training script assumes access to a GPU. If you want to train on a CPU, comment out line 38 in train.py and add --device cpu to the training command.

Model Evaluation

Pedestrian Datasets

To evaluate a trained model, you can execute a version of the following command from within the experiments/pedestrians directory.

python evaluate.py --model <model directory> --checkpoint <epoch number> --data ../processed/<dataset>_test.pkl --output_path results --output_tag <dataset>_<vel if no integration>_12 --node_type PEDESTRIAN

For example, a fully-fleshed out version of this command to evaluate a model without dynamics integration for evaluation on the ETH - University scene would look like:

python evaluate.py --model models/eth_vel --checkpoint 100 --data ../processed/eth_test.pkl --output_path results --output_tag eth_vel_12 --node_type PEDESTRIAN

The same for a model with dynamics integration would look like:

python evaluate.py --model models/eth_attention_radius_3 --checkpoint 100 --data ../processed/eth_test.pkl --output_path results --output_tag eth_12 --node_type PEDESTRIAN

These scripts will produce csv files in the results directory which can then be analyzed in the Result Analysis.ipynb notebook.

nuScenes Dataset

If you just want to use a trained model to generate trajectories and plot them, you can do this in the NuScenes Qualitative.ipynb notebook.

To evaluate a trained model's performance on forecasting vehicles, you can execute a one of the following commands from within the experiments/nuScenes directory.

If you instead wanted to evaluate a trained model's performance on forecasting pedestrians, you can execute a one of the following.

These scripts will produce csv files in the results directory which can then be analyzed in the NuScenes Quantitative.ipynb notebook.

Online Execution

As of December 2020, this repository includes an "online" running capability. In addition to the regular batched mode for training and testing, Trajectron++ can now be executed online on streaming data!

The trajectron/test_online.py script shows how to use it, and can be run as follows (depending on the desired model).

Further, lines 145-151 can be changed to choose different scenes and starting timesteps.

During running, each prediction will be iteratively visualized and saved in a pred_figs/ folder within the specified model folder. For example, if the script loads the int_ee version of Trajectron++ then generated figures will be saved to experiments/nuScenes/models/int_ee/pred_figs/.

Datasets

ETH and UCY Pedestrian Datasets

Preprocessed ETH and UCY datasets are available in this repository, under experiments/pedestrians/raw (e.g., raw/eth/train). The train/validation/test splits are the same as those found in Social GAN.

If you want the original ETH or UCY datasets, you can find them here: ETH Dataset and UCY Dataset.

nuScenes Dataset

If you only want to evaluate models (e.g., produce trajectories and plot them), then the nuScenes mini dataset should be fine. If you want to train a model, then the full nuScenes dataset is required. In either case, you can find them on the dataset website.