Official codes for the paper ''Escaping Stability-Plasticity Dilemma in Online Continual Learning for Motion Forecasting via Synergetic Memory Rehearsal'', Yunlong Lin, Chao Lu, Tongshuai Wu, Xiaocong Zhao, Guodong Du, Yanwei Sun, Zirui Li, and Jianwei Gong.

The experiments in this work are based on INTERACTION dataset.

• The processed data is available in this link for Google Drive.
• Please download the processed data in the direction ./cl_dataset/.

1. System: The codes can be run in Ubuntu 22.04 LTS.
2. Python = 3.9
3. Pytorch = 2.0.0
4. Other required packages are provided in "requirements.txt":

 pip install -r requirements.txt

• Before running codes, please revise root_dir and data_dir in ./utils/args_loading.py to your local paths.
• Parameters for the networks can be also revised in ./utils/args_loading.py.

• --model: the method you want to train and test.
• --buffer_size: the memory size of the continual learning methods to run, and set as 0 when using the vanilla method.
• --dataset: set as "seq-interaction" when continual training, set as "joint-interaction" when joint training.
• --train_task_num: the number of tasks in continual training.
• --debug_mode: True or 1 when you are debugging, only a few batches of samples will be used in each task for a convenient check. False or 0 in the formal training.
• --num_tasks: the number of continual tasks for testing.

After adding the Executable Permissions to the provided bash file (bash_training_and_test.sh), you can directly run the training and testing with command:

./bash_training_and_test_cuda0.sh

│  run_train_test_cuda0.sh
│  test_CL_bi_direct.py
│  test_joint.py
│  test_joint2held_out.py
│  test_single.py
│  train_CL.py
│  train_joint.py
│  train_single.py
│
├─cl_dataset
│  ├─train
│  └─val
├─cl_data_stream
│      joint_dataset.py
│      seq_dataset.py
│      traj_dataset.py
│      traj_joint_dataset.py
│
├─cl_model
│      agem.py
│      continual_model.py
│      jotr.py
│      syrem.py
│      vanilla.py
│      __init__.py
│
├─experiments
│      joint_training.py
│      seq_training_all_task.py
│      testing_1_task_bi_direct.py
│      testing_1_task_joint2held_out.py
│      testing_1_task_single.py
│      testing_joint_tasks.py
│      training_single_task.py
│
├─mapfiles
│      DR_CHN_Merging_ZS0.osm
│      DR_CHN_Merging_ZS0.osm_xy
│      DR_CHN_Merging_ZS2.osm
│      DR_CHN_Merging_ZS2.osm_xy
│      DR_CHN_Roundabout_LN.osm
│      DR_CHN_Roundabout_LN.osm_xy
│      ...
│
├─results
│  ├─logs
│  └─weights
├─traj_predictor
│      baselayers.py
│      decoder.py
│      encoder.py
│      evaluation.py
│      inference.py
│      interaction_model.py
│      losses.py
│      utils.py
│      __init__.py
│
├─utils
│      args_loading.py
│      memory_buffer.py
│      metrics.py
│
└─visualization_utils
        dictionary.py
        dict_utils.py
        extract_original_tv_info.py
        map_vis_without_lanelet.py
        traj_legend.py
        traj_plot.py
        traj_plot_single_fig.py

• Continual training in sequential tasks :

  python train_CL.py
  

• Train the model within a specific task (optional task should be chosen in hyper parameter, refering to args_loading.py):

  python train_single.py
  

• Joint training using multiple datasets (mixed training data among multi-tasks):

  python train_joint.py
  

• Test the model in all tasks (previously learned, current, unseen ones):

python test_CL_bi_direct.py

• Test the model in a specific test set of one task:

python test_single.py

• Test the model in a joint testing set, which covers test data from multiple tasks:

python test_joint.py

⚠ Please set your chosen hyper-parameters (the .sh file in this repo is referred to as an example of setting neccessary hyper-parameters) in args_loading.py before running.

Please feel free to contact our main contributors if you have any questions or suggestions!

Yunlong Lin (Email: [email protected])