DAVA: Disentangling Adversarial Variational Autoencoder

Authors: Benjamin Estermann, Roger Wattenhofer

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare DAVA to models with optimal hyperparameters. Without any hyperparameter tuning, DAVA is competitive on a diverse range of commonly used datasets.
Researcher Affiliation Academia Benjamin Estermann ETH Zürich Switzerland estermann@ethz.ch Roger Wattenhofer ETH Zürich Switzerland wattenhofer@ethz.ch
Pseudocode Yes Algorithm 1 Training DAVA Algorithm 2 update C
Open Source Code Yes The code is available at github.com/besterma/dava.
Open Datasets Yes We namely consider Shapes3D (Burgess & Kim, 2018), Abstract DSprites (Van Steenkiste et al., 2019) and Mpi3d Toy (Gondal et al., 2019). All datasets are part of disentanglement-lib (Locatello et al., 2019b), which we used to train the models we evaluated our metric on.
Dataset Splits No The paper mentions training steps, random seeds, and hyperparameter ranges (Appendix C.1), but does not explicitly state the train/validation/test dataset splits (e.g., percentages or counts) for the datasets used.
Hardware Specification No The paper describes model architectures and hyperparameters used for training and evaluation but does not specify any hardware details like CPU, GPU models, or cloud computing resources.
Software Dependencies No The paper mentions using 'disentanglement-lib' and 'Adam' optimizer, but does not provide specific version numbers for these or other software dependencies like Python or PyTorch.
Experiment Setup Yes The exact hyperparameters considered can be found in appendix C.1. The training of DAVA does not rely on such dataset-specific regularization, therefore all hyperparameters for the training of DAVA were kept constant across all considered datasets. All other hyperparameters closely follow (Locatello et al., 2019b) and were kept consistent across all architectures (baseline and DAVA) and are also reported in appendix C.1.