Parallel and Flexible Sampling from Autoregressive Models via Langevin Dynamics

Authors: Vivek Jayaram, John Thickstun

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present qualitative and quantitive results of Pn F sampling for Wave Net models of audio (van den Oord et al., 2016a) and a Pixel CNN++ model of images (Salimans et al., 2017). In Section 4.2 we show that Pn F sampling can produce samples of comparable quality to ancestral sampling. In Section 4.3 we show that stochastic Pn F sampling is faster than ancestral sampling, when parallelized across a modest number of devices.
Researcher Affiliation Academia Vivek Jayaram * 1 John Thickstun * 1 1Department of Computer Science, University of Washington. Correspondence to: Vivek Jayaram <vjayaram@cs.washington.edu>, John Thickstun <thickstn@cs.washington.edu>.
Pseudocode Yes Algorithm 1 Parallel and Flexible Sampling and Algorithm 2 Stochastic Parallel and Flexible Sampling
Open Source Code Yes Code and examples of Pn F sampling are available at: https://grail.cs.washington.edu/projects/ pnf-sampling/.
Open Datasets Yes For audio experiments we use the VCTK dataset (Veaux et al., 2016) consisting of 44 hours of speech, as well as the Supra Piano dataset (Shi et al., 2019) consisting of 52 hours of piano recordings. For image experiments we use the CIFAR-10 dataset (Krizhevsky, 2009) with the standard train-test split.
Dataset Splits No We use a random 80-20 train-test split of VCTK speakers and piano recordings for evaluation. For image experiments we use the CIFAR-10 dataset (Krizhevsky, 2009) with the standard train-test split. While train and test splits are mentioned, an explicit validation split is not described.
Hardware Specification Yes This behavior is demonstrated in Figure 3 for spectrogram-conditioned Wave Net stochastic Pn F sampling using a cluster of 8 Nvidia Titan Xp GPU s and T = 256.
Software Dependencies No The paper mentions models like Wave Net and Pixel CNN++, but does not specify software dependencies with version numbers.
Experiment Setup No The paper states 'Additional training and hyperparameter details can be found in the appendix.' Since the appendix is not provided, and no specific hyperparameters like learning rate, batch size, or optimizer settings are present in the main text, the details are not explicitly provided here.