Normalizing flow neural networks by JKO scheme

Authors: Chen Xu, Xiuyuan Cheng, Yao Xie

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments with synthetic and real data show that the proposed JKO-i Flow network achieves competitive performance compared with existing flow and diffusion models at a significantly reduced computational and memory cost.
Researcher Affiliation Academia Chen Xu School of Industrial and Systems Engineering Georgia Tech Xiuyuan Cheng Department of Mathematics Duke University Yao Xie School of Industrial and Systems Engineering Georgia Tech
Pseudocode Yes Algorithm 1 Block-wise JKO-i Flow training
Open Source Code Yes Codes are available at https://github.com/hamrel-cxu/JKO-i Flow.
Open Datasets Yes The four high-dimensional real datasets (POWER, GAS, MINIBOONE, BSDS300) come from the University of California Irvine (UCI) machine learning data repository, and we follow the pre-processing procedures of [Papamakarios et al., 2017].
Dataset Splits No POWER: 1.85M training sample and 205K test sample. GAS: 1M training sample, 100K test sample. MINIBOONE: 32K training sample, 3.6K test sample. BSDS300: 1.05M training sample, 250K test sample. (No explicit mention of validation set size or split.)
Hardware Specification Yes Specifically, on a single A100 GPU, our experiments took 24 hours on CIFAR10 and 30 hours on Imagenet-32.
Software Dependencies No All experiments are conducted using Py Torch [Paszke et al., 2019] and Py Torch Geometric [Fey and Lenssen, 2019]. (Specific version numbers for PyTorch and PyTorch Geometric are not stated.)
Experiment Setup Yes The initial schedule is by setting h0 = 0.75, ρ = 1.2. We set hmax = 5 for rose, checkerboard, and Olympic rings, and hmax = 3 for fractal tree. For reparametrization and refinement, we use η = 0.5 in the reparameterization iterations for all examples. The learning rate is 5e-3. On MNIST, we fix the batch size to be 2000 during training, and we train 15K batches per JKO-i Flow block. We fix the learning rate to be 1e-3.