Improved Techniques for Training Score-Based Generative Models

Authors: Yang Song, Stefano Ermon

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide a new theoretical analysis of learning and sampling from score-based models in high dimensional spaces, explaining existing failure modes and motivating new solutions that generalize across datasets. To enhance stability, we also propose to maintain an exponential moving average of model weights. With these improvements, we can scale scorebased generative models to various image datasets, with diverse resolutions ranging from 64 64 to 256 256. Our score-based models can generate high-fidelity samples that rival best-in-class GANs on various image datasets, including Celeb A, FFHQ, and several LSUN categories.
Researcher Affiliation Academia Yang Song Computer Science Department Stanford University yangsong@cs.stanford.edu Stefano Ermon Computer Science Department Stanford University ermon@cs.stanford.edu
Pseudocode Yes Algorithm 1 Annealed Langevin dynamics [1]
Open Source Code No No explicit statement about releasing the source code for the methodology described in this paper or a link to a code repository was found.
Open Datasets Yes Our score-based models can generate high-fidelity samples that rival best-in-class GANs on various image datasets, including Celeb A, FFHQ, and several LSUN categories. (with citations [2], [16], [27] pointing to CIFAR-10, Celeb A, and LSUN datasets respectively)
Dataset Splits No The paper does not explicitly provide training/test/validation dataset splits, nor does it describe cross-validation. It mentions training and test data, but no explicit validation split.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, or memory amounts) used for running the experiments were provided. The paper does not specify the computational resources.
Software Dependencies No No specific ancillary software details with version numbers (e.g., library names with versions) needed to replicate the experiment were provided. The paper does not list software dependencies with their versions.
Experiment Setup Yes For a complete description on experimental details and more results, please refer to Appendix B and C. (From Appendix B.1: The initial learning rate is 0.0001, which decays by 0.9999 for every 10000 steps. We use Adam [26] optimizer with β1 = 0.9, β2 = 0.999 and ϵ = 10 8. The batch size is 128. We train the NCSN models for 1000000 iterations for Celeb A 64 64 and CIFAR-10 32 32.)