MMVAE+: Enhancing the Generative Quality of Multimodal VAEs without Compromises

Authors: Emanuele Palumbo, Imant Daunhawer, Julia E Vogt

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

Reproducibility Variable Result LLM Response
Research Type Experimental We report experimental results for two challenging datasets introduced in previous work, namely Poly MNIST (Sutter et al., 2021) and Caltech Birds (CUB) Image-Captions (Shi et al., 2019; Wah et al., 2011).
Researcher Affiliation Academia 1 ETH AI Center, Z urich, Switzerland 2 Department of Computer Science, ETH Z urich, Switzerland
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes We provide details about datasets, metrics, hyperparameters and architectures in the Appendix, which can aid reproducibility of our method along with the provided code.
Open Datasets Yes We report experimental results for two challenging datasets introduced in previous work, namely Poly MNIST (Sutter et al., 2021) and Caltech Birds (CUB) Image-Captions (Shi et al., 2019; Wah et al., 2011).
Dataset Splits No The paper mentions using a 'test set' for evaluation but does not specify the train/validation/test splits, percentages, or methodology for creating these splits.
Hardware Specification No The paper does not specify the hardware used for running the experiments.
Software Dependencies Yes To compute FID scores we use the implementation from Seitzer (2020).
Experiment Setup Yes We compare all models for a range of representative values of the regularization hyperparameter β {1.0, 2.5, 5.0}. ... All models are trained using the Adam optimizer (Kingma & Ba, 2014). We choose a learning rate of 5e-4 for MVAE and Mo Po E-VAE, following Daunhawer et al. (2022), while for MMVAE and MMVAE+ we choose 1e-3.