A Langevin-like Sampler for Discrete Distributions

Authors: Ruqi Zhang, Xingchao Liu, Qiang Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We provide extensive experimental results, including Ising models, restricted Boltzmann machines, deep energy-based models, binary Bayesian neural networks and text generation, to demonstrate the superiority of DLP in general settings.
Researcher Affiliation Academia 1The University of Texas at Austin. Correspondence to: Ruqi Zhang <ruqiz@utexas.edu>.
Pseudocode Yes Algorithm 1 Samplers with Discrete Langevin Proposal (DULA and DMALA).
Open Source Code Yes We released the code at https://github. com/ruqizhang/discrete-langevin.
Open Datasets Yes Following Grathwohl et al. (2021), we train {W, a, b} with contrastive divergence (Hinton, 2002) on the MNIST dataset for one epoch. and We conduct regression on four UCI datasets (Dua & Graff, 2017).
Dataset Splits Yes The reported results are from the model which performs the best on the validation set.
Hardware Specification No The paper mentions that computations can be done "in parallel on CPUs and GPUs" but does not provide specific hardware details such as GPU/CPU models or memory specifications used for experiments.
Software Dependencies No The paper states "All methods are implemented in Pytorch" but does not provide specific version numbers for PyTorch or any other software dependencies needed for replication.
Experiment Setup Yes DMALA and DULA use a stepsize α of 0.4 and 0.2 respectively. and We set the batch size to 100. The stepsize α is set to be 0.1 for DULA and 0.2 for DMALA, respectively. The model is optimized by an Adam optimizer with a learning rate of 0.001.