An Identifiable Double VAE For Disentangled Representations

Authors: Graziano Mita, Maurizio Filippone, Pietro Michiardi

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results indicate superior performance with respect to state-of-the-art approaches, according to several established metrics proposed in the literature on disentanglement.
Researcher Affiliation Collaboration 1EURECOM, 06410 Biot (France) 2SAP Labs France, 06250 Mougins (France).
Pseudocode No The paper describes the algorithms and models in text and mathematical formulas but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions that 'All methods have been implemented in Py Torch (Paszke et al., 2019)' and that 'The implementation of the metrics is based on Locatello et al. (2019)', but it does not provide an explicit statement or link for the source code of their proposed method (IDVAE).
Open Datasets Yes We consider four common datasets in the disentanglement literature, where observations are images built as a deterministic function of known generative factors: DSPRITES (Higgins et al., 2017), SHAPES3D (Kim & Mnih, 2018), CARS3D (Reed et al., 2015) and SMALLNORB (Le Cun et al., 2004).
Dataset Splits No The paper does not provide specific numerical train/validation/test splits (e.g., percentages or counts) or explicitly reference predefined splits that include validation data. It mentions '300 000 training iterations' and evaluation 'At testing time', implying training and test sets, but no explicit validation split.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No All methods have been implemented in Py Torch (Paszke et al., 2019).
Experiment Setup Yes We tried six different values of regularization strength associated to the target regularization term of each method β for β-VAE, IVAE and IDVAE, and γ for FULLVAE: [1, 2, 4, 6, 8, 16]. For each model configuration and dataset, we run the training procedure with 10 random seeds, given that all methods are susceptible to initialization values. After 300 000 training iterations, every model is evaluated according to the disentanglement metrics described above. The latent dimension z is fixed to the true number of ground-truth factors.