Home

Awesome

PWC

Transformer Networks for Trajectory Forecasting

This is the code for the paper <a href="https://arxiv.org/abs/2003.08111">Transformer Networks for Trajectory Forecasting</a>

Requirements

Usage

Data setup

The dataset folder must have the following structure:

- dataset
  - dataset_name
    - train_folder
    - test_folder
    - validation_folder (optional) 
    - clusters.mat (For quantizedTF)

Individual Transformer

To train just run the train_individual.py with different parameters

example: to train on the data for eth

CUDA_VISIBLE_DEVICES=0 python train_individualTF.py --dataset_name eth --name eth --max_epoch 240 --batch_size 100 --name eth_train --factor 1

QuantizedTF

Step1: Create the clusters

NOTE: We used a pytorch based method that use GPUs to lower the computational time, but it requires both a GPU and a high amount of RAM (25 GB).
Since clusters do not change over time they can be created with any code, you just need to create a file with the centroids inside the dataset/dataset_name folder

For ease of use the cluster informations are already upladed for eth+ucy

To create the cluster_mat file run kmeans.py

CUDA_VISIBLE_DEVICES=0 python kmeans.py --dataset_name eth

After that put the clusters.mat inside the appropriate dataset folder.

Step 2: Train the quantized

Run ClassifyTF.py

CUDA_VISIBLE_DEVICES=0 python train_quantizedTF.py --dataset_name zara1 --name zara1 --batch_size 1024

Step 3: Evaluate Best-of-N

Run test_class.py with the parameters for the dataset_name, the name of the trained model, the epoch to test and the number of samples

CUDA_VISIBLE_DEVICES=0 python test_quantizedTF.py --dataset_name eth --name eth --batch_size 1024 --epoch 00030 --num_samples 20

Visualization

The training loss, validation loss, mad and fad for the test can be seen for each epoch by running tensorboard

tensorboard --logdir logs

Citation

If you use the code please cite our paper.

@misc{giuliari2020transformer,
    title={Transformer Networks for Trajectory Forecasting},
    author={Francesco Giuliari and Irtiza Hasan and Marco Cristani and Fabio Galasso},
    year={2020},
    eprint={2003.08111},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Thanks

TODO

Changelog