TANGO

Data Generation

To generate the simulated datasets, go to the data directory first

cd data

where we have a separate script for generating each dataset. Hyperparameters used for datasets evaluated in the paper are saved as default values for each argument. To generate the simple spring dataset, run

python generate_dataset.py

To generate the forced spring dataset, run

python generate_dataset_external.py

To generate the damped spring dataset, run

python generate_dataset_damped.py

To generate the pendulum dataset, run

python generate_dataset_pendulum.py

Training

For all training procedures, use "--n-balls 5" for the spring datasets and use "--n-balls 3" for the pendulum dataset. We use the --data option to distinguish the dataset to be trained on. Supported dataset types include simple_spring, damped_spring, forced_spring, pendulum.

Training TANGO and LGODE

To start the training of TANGO and LGODE, use run_model.py. To run LGODE, use the command

python run_model.py --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 

with appropriate arguments. To run TANGO, provide a non-negative value for the --reverse_f_lambda option

python run_model.py --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 --reverse_f_lambda 0.5

To run the ablation using the time-reversal loss following the original definition of time-reversal, use the --use_trsode option

python run_model.py --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 --reverse_f_lambda 0.5 --use_trsode

To run the ablation using the difference between groud truth and backward trajectories, use the --reverse_gt_lambda option instead

python run_model.py --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 --reverse_gt_lambda 0.5

Training HODEN, TRS-ODEN, and TRS-ODEN_GNN

To train these models, use the run_models_trsode.py script, and use the --function_type option to specify the model.

Running HODEN

python run_models_trsode.py --function_type hamiltonian --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 --reverse_f_lambda 0.5

Running TRS-ODEN

python run_models_trsode.py --function_type ode --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 --reverse_f_lambda 0.5

Running TRS-ODEN_GNN

CUDA_VISIBLE_DEVICES=0 python run_models_trsode.py --function_type gnn --data simple_spring --n-balls 5 --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 --reverse_f_lambda 0.5

Training LatentODE

To train LatentODE, go to latent_ode folder and then run

python run_model.py --data simple_spring --n-balls 5 --latent-ode --train_cut 20000 --test_cut 5000 --sample-percent-train 0.6 --sample-percent-train 0.6 -l 15 -u 1000 -g 50 --rec-layers 4 --gen-layers 2 --rec-dim 100 --lr 1e-3