Train VAE
python src/train_VAE.py tools/train_vae_cifar10.json
Train VAEBM
Specify the checkpoint argument in train_vaebm_cifar10.json, and then run
python src/train_VAEBM.py tools/train_vaebm_cifar10.json
python -m pytorch_fid /data/10707project/output/groundtruth /data/10707project/output/generated/beta_1/random --device cuda:1 python -m pytorch_fid /data/10707project/output/groundtruth /data/10707project/output/generated/beta_2/mulog_added --device cuda:1 python -m pytorch_fid /data/10707project/output/groundtruth /data/10707project/output/generated/beta_2/mulog_mcmc --device cuda:1
compare fid of randomly generated images between VAE and VAEBM/EBM
python -m pytorch_fid /data/10707project/output/groundtruth /data/10707project/output/vae/celeba64/beta_1/random --device cuda:0
python -m pytorch_fid /data/10707project/output/groundtruth /data/10707project/output/vae/celeba64/beta_1/recon --device cuda:0
random samples VAE VAEBM beta=1 75.571 75.591 beta=2 79.821 73.580 beta=4 88.479 83.063 beta=10 107.176 96.277
reconstructions VAE VAEBM (mulog added) mulog mcmc mulog both mcmc beta=1 50.408 50.790 52.668 43.855 beta=2 60.915 60.978 58.621 53.634 beta=4 74.754 75.061 71.409 61.221 beta=10 98.109 98.484 90.154 85.666