Diffusion bridges vector quantized variational autoencoders

Authors: Max Cohen, Guillaume Quispe, Sylvain Le Corff, Charles Ollion, Eric Moulines

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that our model is competitive with the autoregressive prior on the mini-Imagenet and CIFAR dataset and is efficient in both optimization and sampling. We benchmarked our model using three metrics, in order to highlight the performances of the proposed prior, the quality of produced samples as well as the associated computation costs. Results are given as a comparison to the original Pixel CNN prior for both the mini Image Net (see Table 2) and the CIFAR10 (see Table 3) datasets.
Researcher Affiliation Collaboration 1Samovar, T el ecom Sud Paris, D epartement CITI, Institut Polytechnique de Paris, Palaiseau, France. 2Oze Energies, Charenton Le-Pont, France. 3Centre de Math ematiques Appliqu ees, Ecole polytechnique, Institut Polytechnique de Paris, Palaiseau, France.
Pseudocode Yes Algorithm 1 Training procedure Algorithm 2 Sampling procedure
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets Yes CIFAR10. The CIFAR dataset consists of inputs x of dimensions 32 32 with 3 channels. mini Image Net. mini Image Net was introduced by (Vinyals et al., 2016) to offer more complexity than CIFAR10...
Dataset Splits No The paper mentions using a validation dataset (e.g., 'Metrics are computed on the validation dataset.'), but it does not specify explicit split percentages, absolute sample counts, or detailed methodology for dataset partitioning to enable reproduction.
Hardware Specification Yes We evaluated the computation cost of sampling a batch of 32 images, on a GTX TITAN Xp GPU card.
Software Dependencies No The paper mentions using a 'U-net like architecture' but does not provide specific software dependencies or library versions (e.g., PyTorch 1.x, TensorFlow 2.x) that were used for implementation.
Experiment Setup Yes Concerning our diffusion prior, we choose the Ornstein-Uhlenbeck process setting η = 2, z = 0 and ϑ = 1, with T = 1000. In practice τ = 1 and a simple schedule from 10 to 0.1 for τt was considered in this work.