This code supports the ICLR 2023 Submission DAVA: Disentangling Adversarial Variational Autoencoder

We provide two environment files, dava.yml and pipe.yml. Due to versioning conflicts with disentanglement-lib, we suggest to use separate environments for training DAVA and evaluating the PIPE metric. To our knowledge, for these two environments, no issues should arise when evaluating a model in a different environment than it was trained. We provide dsprites as an example dataset in the correct format.

To reproduce DAVA: python dava/train_dava.py --dataset_path=datasets/dsprites.npz --z_dim=10 --batch_size=128 --num_steps=150000 --num_channels=1 --store_path=~/dava/ --disc_weight=0.3 --c_max=8 --iteration_threshold=12800000 --disc_instance_norm --gamma=500 --random_seed=0 --max_grad_norm=1 --learning_rate=0.0001 --disc_max_grad_norm=1 --disc_learning_rate=0.0001 --use_mixed_batches --model_scale_factor=1 --start_kl_step=0 --disc_confidence_value=0.6 --disc_decoder_weight=0.001

To reproduce any of the other models, for example FactorVAE: python dava/train_factor_vae.py --batch_size=128 --num_channels=1 --learning_rate=0.0001 --dataset_path=datasets/dsprites.npz --store_path=~/factor_vae/ --gamma=5 --num_steps=150000 --random_seed=0 --max_grad_norm=1

Reproducing the results of the PIPE metric is a bit more intricate. The code for computing the PIPE metric is provided in pipe_metric.py Models were trained using disentanblement-lib. We used dlib_reproduce to train Beta-TCVAEs and FactorVAEs with 6 hyperparameters and 5 random seeds each. We did this for each dataset considered. We then used disentanglement-lib to compute all supervised as well as the UDR metric. We extended disentanglement-lib so that with slight changes, it can be used to evalaute PIPE as well. We provide the necessary files here. pipe_metric.py needs to be placed in disentanglement_lib.evaluation.metrics. evaluate.py replaces disentanglement_lib.evaluation.evaluate.py. compute_metrics_vae.py gives an overview on how to start the metric computation. This works for both models reproduced with disentanglement-lib and models trained in the DAVA environment.

We further used disentanglement-lib to reproduce the results of the abstract reasoning study for BetaTCVAE and FactorVAE. Models there could be evaluated using the same procedure as explained before.