keepflow

This repository aims to provide the code base for temporal several prediction tasks, such as trajectory :curly_loop:, motion :dancer:, and video :movie_camera:.

News

• Jan 2024: repository public on GitHub :confetti_ball::confetti_ball::confetti_ball:

Dependencies

pip install -r requirements.txt
git submodule init
git submodule update

Data preprocessing

Trajectory prediction

datasets supported by trajdata such as ETH/UCY and Stanford Drone Datasets.

Please follow the trajdata instruction and specify the dataset path by cfg.DATA.PATH in keepflow/utils/default_params.py.

• ETH/UCY

  cfg.DATA.PATH/traj/raw/all_data
          ├── biwi_eth.txt
          ├── biwi_hotel.txt
          ├── crowds_zara01.txt
          ├── crowds_zara02.txt
          ├── crowds_zara03.txt
          ├── students001.txt
          ├── students003.txt
          └── uni_examples.txt

JackRabbot Dataset and Benchmark Trajectory Forecasting

Please follow the jrdb-traj instruction and execute dataload.sh and preprocess.sh in extern/traj/jrdb-traj/.

Supported Models

Trajectory prediction

• SocialLSTM based on socialGAN GitHub Code
• Trajectron++ based on Trajectron++ GitHub Code
• Motion Indeterminacy Diffusion based on MID GitHub Code
• FlowChain based on FlowChain GitHub Code

Training

python train.py --config_file CONFIG_FILE_PATH --device DEVICE

Example

python train.py --config_file configs/traj/FlowChain/CIF_separate_cond_v_trajectron/eth.yml --device cuda:0

You can execute training consecutively by script

python scripts/run_configs.py --config_dir configs/traj/FlowChain/CIF_separate_cond_v_trajectron/ --device cuda:0 (--test_only)

Evaluation

python test.py --config_file CONFIG_FILE_PATH --device DEVICE (--visualize)

Example

python test.py --config_file configs/traj/FlowChain/CIF_separate_cond_v_trajectron/eth.yml --device cuda:0