Generative Adversarial Symmetry Discovery (LieGAN)

Code for the ICML 2023 paper, Generative Adversarial Symmetry Discovery.

Experiments

N-Body Trajectory

To discover symmetry from the task of 2-body trajectory prediction, run one of the following:

# LieGAN
python main_lagan.py --g_init 2*2_factorization --lamda 1 --sigma_init 1 --num_epochs 100
# Augerino*
python main_augerino.py --g_init 2*2_factorization --lamda 1 --num_epochs 100
# SymmetryGAN
python main_lagan.py --g_init 2*2_factorization --lamda 1 --sigma_init 1 --num_epochs 100 --model sgan

To discover symmetry in a larger search space (Figure 3c), run the following:

python main_lagan.py --g_init 4*4_factorization --lamda 1e1 --sigma_init 3 --num_epochs 100

To run the prediction experiments for EMLP and data augmentation, refer to emlp.ipynb and aug_pred.ipynb. We have provided the discovery results used for these experiments in ./saved_model/results/.

As is stated in Section 5.2 in our paper, we need to slightly modify the EMLP implementation of SVD. It is located in line 307 of emlp/reps/representation.py, where we need to change the threshold from 1e-5 to larger values, like 5e-3.

To discover symmetry from the task of 3-body trajectory prediction:

python main_lagan.py --task traj_pred_3body --dataset_name 3body --g_init 2*2_factorization --lamda 1 --sigma_init 1

Discrete Rotation Invariant Regression

python main_lagan.py --task discrete_rotation_synthetic --dataset_size 20000 --y_type scalar --coef_dist uniform_int_grid --uniform_max 10

Top Tagging

To discover symmetry, first download the top tagging dataset to ./data/top-tagging. We only need train.h5 for symmetry discovery purpose. Then run the following command:

python main_lagan.py --task top_tagging --lamda 1 --g_init random --n_channel 7 --y_type scalar --sigma_init 1 --eta 0.1 --n_component 2

To run the top tagging prediction task, first download the converted dataset from this link to the data directory ./data/top-tagging-converted. Then,

cd LorentzNet

The content of this directory is modified from the codebase of LorentzNet. We mainly made our modifications in the file models.py, where we implemented LieGNN as proposed in our paper. To run the experiments:

# LorentzNet (original)
python -m torch.distributed.launch --nproc_per_node=4 top_tagging.py \
        --batch_size=32 --epochs=35 --warmup_epochs=5 \
        --n_layers=6 --n_hidden=72 --lr=3e-4 \
        --c_weight=0.005 --dropout=0.2 --weight_decay=0.01 \
        --exp_name=lorentznet --datadir ../data/top-tagging-converted
# LieGNN (symmetry discovered by LieGAN)
python -m torch.distributed.launch --nproc_per_node=4 top_tagging.py \
        --batch_size=32 --epochs=35 --warmup_epochs=5 \
        --n_layers=6 --n_hidden=72 --lr=3e-4 \
        --c_weight=0.005 --dropout=0.2 --weight_decay=0.01 \
        --exp_name=liegnn --datadir ../data/top-tagging-converted --metric saved_metric/metric_liegan.pt
# Augerino+GNN (symmetry discovered by Augerino+)
python -m torch.distributed.launch --nproc_per_node=4 top_tagging.py \
        --batch_size=32 --epochs=35 --warmup_epochs=5 \
        --n_layers=6 --n_hidden=72 --lr=3e-4 \
        --c_weight=0.005 --dropout=0.2 --weight_decay=0.01 \
        --exp_name=augerino --datadir ../data/top-tagging-converted --metric saved_metric/metric_augerino.pt

Rotated MNIST

python main_lagan.py --task MNIST --g_init affine --x_type grid --y_type scalar --lamda 0.1 --lr_g 2e-4 --reg_type fourier --num_epochs 15

Cite

@article{yang2023generative,
  title={Generative Adversarial Symmetry Discovery},
  author={Yang, Jianke and Walters, Robin and Dehmamy, Nima and Yu, Rose},
  booktitle={International Conference on Machine Learning},
  organization={PMLR},
  year={2023}
}