Code for the paper Bi-level Score Matching for Learning Energy-based Latent Variable Models.

See environment.yml. You can create the environment by running

conda env create -f environment.yml

To compare BiSM with baselines (including CD, PCD, DSM, SSM, NCE, VNCE) on the toy (checkerboard) dataset, run

python checkerboard_exps.py

To see how N and K influence the results, run

python freyface_exps.py

To train an EBLVM on the mnist dataset (1 * 28 * 28) with BiSM, run

python tune_latent_ebm_resnet_mnist_bimdsm.py

To train an EBLVM on the cifar10 dataset (3 * 32 * 32) with BiSM, run

python tune_latent_ebm_resnet_cifar10_bimdsm.py

To train an EBLVM on the celeba dataset (3 * 32 * 32) with BiSM, run

python tune_latent_ebm_resnet_celeba_bimdsm.py

To compare BiSM with baselines (including CD, PCD, DSM, SSM, NCE, VNCE) on the toy (checkerboard) dataset, run

python checkerboard_exps.py

To train an EBM on the mnist dataset (1 * 28 * 28) with MDSM, run

python tune_ebm_resnet_mnist_mdsm.py

To train an EBM on the cifar10 dataset (3 * 32 * 32) with MDSM, run

python tune_ebm_resnet_cifar10_mdsm.py

• The code will detect free GPUs by command gpustat and run on these GPUs. You can manually assign GPUs by modify the devices argument of function task_assign in the above .py files.

• The downloaded dataset and running result will be saved to workspace directory by default.