Official PyTorch implementation for "Understanding Instance-based Interpretability of Variational Auto-Encoders".
This repo implements VAE-TracIn, an instance based interpretation method for variational autoencoders. See paper via this link.
To train VAE models, cd ./Beta_VAE/, modify dataset paths in paths.py, and then run
- MNIST:
python main.py --dataset mnist --seed 1 --optim sgd --lr 1e-4 --objective H --model MNIST --batch_size 64 --z_dim 128 --max_iter 1.5e6 --beta 4 --viz_on False --viz_name mnist_beta4_z128_sgd - CIFAR subclass:
python main.py --dataset cifar_zero --seed 1 --optim sgd --lr 3e-4 --objective H --model H --batch_size 64 --z_dim 64 --max_iter 1.5e6 --beta 2 --viz_on False --viz_name cifar_zero_beta2_z64_sgd - CIFAR:
python main.py --dataset cifar --seed 1 --optim sgd --lr 3e-4 --objective H --model H --batch_size 64 --z_dim 128 --max_iter 1.5e6 --beta 2 --viz_on False --viz_name cifar_beta2_z128_sgd - MNIST data cleaning:
python main.py --dataset mnist_emnist_extra --seed 1 --optim sgd --lr 1e-4 --objective H --model MNIST --batch_size 64 --z_dim 128 --max_iter 1.5e6 --beta 4 --viz_on False --viz_name mnist_emnist_extra_beta4_z128_sgd - CIFAR data cleaning:
python main.py --dataset cifar_celeba_extra --seed 1 --optim sgd --lr 3e-4 --objective H --model H --batch_size 64 --z_dim 128 --max_iter 1.5e6 --beta 2 --viz_on False --viz_name cifar_celeba_extra_beta2_z128_sgd
Note: in CIFAR subclass experiments, the dataset parameter can be cifar_zero through cifar_nine. The cifar_zero pretrained models are provided.
- MNIST:
python VAE_TracIn_testinf.py --beta 4 --latent_dim 128 --dataset mnist --task train - CIFAR:
python VAE_TracIn_testinf.py --beta 2 --latent_dim 128 --dataset cifar --task train - CIFAR subclass:
python VAE_TracIn_testinf.py --beta 2 --latent_dim 64 --dataset cifar_zero --task train
- MNIST:
python VAE_TracIn_selfinf.py --beta 4 --latent_dim 128 --dataset mnist --last_epoch - CIFAR:
python VAE_TracIn_selfinf.py --beta 2 --latent_dim 128 --dataset cifar --last_epoch - CIFAR subclass:
python VAE_TracIn_selfinf.py --beta 2 --latent_dim 64 --dataset cifar_zero --last_epoch - MNIST data cleaning:
python VAE_TracIn_selfinf.py --beta 4 --latent_dim 128 --dataset mnist_emnist_extra --last_epoch - CIFAR data cleaning:
python VAE_TracIn_selfinf.py --beta 2 --latent_dim 128 --dataset cifar_celeba_extra --last_epoch
- MNIST:
python VAE_TracIn_testinf.py --beta 4 --latent_dim 128 --dataset mnist --task test --load_from_jpg --load_from_jpg_dir output/images/test_mnist - CIFAR:
python VAE_TracIn_testinf.py --beta 2 --latent_dim 128 --dataset cifar --task test --load_from_jpg --load_from_jpg_dir output/images/test_cifar - CIFAR subclass:
python VAE_TracIn_testinf.py --beta 2 --latent_dim 64 --dataset cifar_zero --task test --load_from_jpg --load_from_jpg_dir output/images/test_cifar_zero