Wavelet Score-Based Generative Modeling

Authors: Florentin Guth, Simon Coste, Valentin De Bortoli, Stephane Mallat

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 4, we present numerical results on Gaussian distributions, the φ4 physical model at phase transition, and the Celeb A-HQ image dataset [19]. and Theorems controlling errors of time discretizations of SGMs, proving accelerations obtained by scale separation with wavelets. These results are empirically verified by showing that WSGM provides an acceleration for the synthesis of physical processes at phase transition and natural image datasets.
Researcher Affiliation Collaboration Florentin Guth Computer Science Department, ENS, CNRS, PSL University Simon Coste Computer Science Department, ENS, CNRS, PSL University Valentin De Bortoli Computer Science Department, ENS, CNRS, PSL University Stéphane Mallat Collège de France, Paris, France Flatiron Institute, New York, USA
Pseudocode Yes The generative process is illustrated in Figure 1 and its pseudocode is given in Algorithm 1 in Appendix S2.
Open Source Code Yes All the code needed to reproduce the experiments is provided in the supplemental material.
Open Datasets Yes We present numerical results on Gaussian distributions, the φ4 physical model at phase transition, and the Celeb A-HQ image dataset [19].
Dataset Splits No The paper mentions training on Celeb A-HQ but does not explicitly state the training, validation, and test dataset splits within the main text.
Hardware Specification Yes We estimate the total amount of compute used during the preparation of this paper, including preliminary experiments, at around 10k hours of NVIDIA A100 GPUs.
Software Dependencies No The paper mentions general software components like 'neural network' and 'U-Net architecture' but does not provide specific software dependencies with version numbers.
Experiment Setup Yes Following [38], the global scores sθ(x) are parameterized by a neural network with a U-Net architecture. It has 3 residual blocks at each scale, and includes multi-head attention layers at lower scales. The conditional scores sθ( xj|xj) are parameterized in the same way, and the conditioning on the low frequencies xj is done with a simple input concatenation along channels [38, 40]. The details of the architecture are in Appendix S9. We use a uniform discretization of the backward SDE to stay in the setting of Theorem 2, and show that WSGM still obtains satisfactory results in this case.