jeongwhanchoi / LEAP

Learnable Path in Neural Controlled Differential Equations (AAAI2023)

Learnable Path in Neural Controlled Differential Equations [arXiv]

The overall architecture of LEAP

create conda environments

install conda environments

conda env create --file leap.yml 

train LEAP

conda activate leap

move to experiments folder

cd experiments/

running shell file at /AAAI_LEAP/experiments/

run code mujoco_0.3_final.sh

sh mujoco_0.3_final.sh 

BASELINES MUJOCO

python3 -u mujoco.py --h_channels 80 --hh_channels 50 --layers 5 --lr 0.0001 --c1 0.0001 --c2 0.0001 --method "rk4" --weight_decay 0  --missing_rate 0.3 --time_seq 50 --y_seq 10 --intensity '' --epoch 2 --model ncde_forecasting
python3 -u mujoco.py --h_channels 80 --hh_channels 50 --layers 5 --lr 0.0001 --c1 0.0001 --c2 0.0001 --method "rk4" --weight_decay 0  --missing_rate 0.3 --time_seq 50 --y_seq 10 --intensity 'True' --epoch 2 --model decay_forecasting
python3 -u mujoco.py --h_channels 80 --hh_channels 50 --layers 5 --lr 0.0001 --c1 0.0001 --c2 0.0001 --method "rk4" --weight_decay 0  --missing_rate 0.3 --time_seq 50 --y_seq 10 --intensity 'True' --epoch 2 --model odernn_forecasting
python3 -u mujoco.py --h_channels 80 --hh_channels 50 --layers 5 --lr 0.0001 --c1 0.0001 --c2 0.0001 --method "rk4" --weight_decay 0  --missing_rate 0.3 --time_seq 50 --y_seq 10 --intensity 'True' --epoch 2 --model dt_forecasting
python3 -u mujoco.py --h_channels 80 --hh_channels 50 --layers 5 --lr 0.0001 --c1 0.0001 --c2 0.0001 --method "rk4" --weight_decay 0  --missing_rate 0.3 --time_seq 50 --y_seq 10 --intensity '' --epoch 2 --model gruode_forecasting

run code uea.sh

sh uea.sh 

BASELINES UEA

python3 -u uea.py --dataset_name CharacterTrajectories --h_channels 40 --hh_channels 100 --layer 3 --lr 0.001 --c1 1e-6 --c2 0 --method "rk4" --weight_decay 0  --missing_rate 0.3 --model='ncde'
python3 -u uea.py --dataset_name CharacterTrajectories --h_channels 40 --hh_channels 100 --layer 3 --lr 0.001 --c1 1e-6 --c2 0 --method "rk4" --weight_decay 0  --missing_rate 0.3 --model='gruode'

BASELINES UEA

NEEDS data preprocessing. (Due to the submission memory limitation)

When you run this code, it would start from preprocessing step (it may take a while.).

python3 -u uea.py --dataset_name CharacterTrajectories --h_channels 40 --hh_channels 100 --layer 3 --lr 0.001 --c1 1e-6 --c2 0 --method "rk4" --weight_decay 0  --intensity 'True' --missing_rate 0.3 --model='dt'
python3 -u uea.py --dataset_name CharacterTrajectories --h_channels 40 --hh_channels 100 --layer 3 --lr 0.001 --c1 1e-6 --c2 0 --method "rk4" --weight_decay 0  --intensity 'True' --missing_rate 0.3 --model='decay'
python3 -u uea.py --dataset_name CharacterTrajectories --h_channels 40 --hh_channels 100 --layer 3 --lr 0.001 --c1 1e-6 --c2 0 --method "rk4" --weight_decay 0  --intensity 'True' --missing_rate 0.3 --model='odernn'

Citation

@article{jhin2023learnable,
  title={Learnable Path in Neural Controlled Differential Equations},
  author={Jhin, Sheo Yon and Jo, Minju and Kook, Seungji and Park, Noseong and Woo, Sungpil and Lim, Sunhwan},
  journal={AAAI},
  year={2023}
}