BIVA: A Very Deep Hierarchy of Latent Variables for Generative Modeling

Authors: Lars Maaløe, Marco Fraccaro, Valentin Liévin, Ole Winther

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental BIVA is empirically evaluated by (i) an ablation study analyzing each novel component, (ii) likelihood and semi-supervised classification results on binary images, (iii) likelihood results on natural images, and (iv) an analysis of anomaly detection in complex data distributions.
Researcher Affiliation Collaboration Lars Maaløe Corti Copenhagen Denmark lm@corti.ai Marco Fraccaro Unumed Copenhagen Denmark mf@unumed.com Valentin Liévin & Ole Winther Technical University of Denmark Copenhagen Denmark {valv,olwi}@dtu.dk
Pseudocode No No pseudocode or algorithm blocks were found.
Open Source Code Yes Source code (Tensorflow): https://github.com/larsmaaloee/BIVA. Source code (Py Torch): https://github.com/vlievin/biva-pytorch.
Open Datasets Yes We evaluate BIVA L = 6 in terms of test log-likelihood on statically binarized MNIST [43], dynamically binarized MNIST [28] and dynamically binarized OMNIGLOT [25]. We trained and evaluated BIVA L = 15 on 32x32 CIFAR-10, 32x32 Image Net [57], and another BIVA L = 20 on 64x64 Celeb A [27].
Dataset Splits No The paper does not explicitly state the training/validation/test dataset splits with specific percentages or counts. It mentions 'test set' for evaluation but not the full split methodology for training and validation.
Hardware Specification No The paper does not specify any hardware details like GPU/CPU models or specific machine configurations used for running experiments.
Software Dependencies No The paper mentions 'Source code (Tensorflow): https://github.com/larsmaaloee/BIVA. Source code (Py Torch): https://github.com/vlievin/biva-pytorch.' but does not specify version numbers for these software dependencies or any other libraries.
Experiment Setup Yes We employ a free bits strategy with λ = 2 [23] for all experiments to avoid latent variable collapse during the initial training epochs. Trained models are reported with 1 importance weighted sample, L1, and 1000 importance weighted samples, L1e3 [3]. We evaluate the natural image experiments by bits per dimension (bits/dim), L/(hwc log(2)), where h, w, c denote the height, width, and channels respectively. For a detailed description of the experimental setup see Appendix C and the source code.