Building Normalizing Flows with Stochastic Interpolants
Authors: Michael Samuel Albergo, Eric Vanden-Eijnden
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Benchmarking on density estimation tasks illustrates that the learned flow can match and surpass conventional continuous flows at a fraction of the cost, and compares well with diffusions on image generation on CIFAR-10 and Image Net 32 32. The method scales ab-initio ODE flows to previously unreachable image resolutions, demonstrated up to 128 128. |
| Researcher Affiliation | Academia | Michael S. Albergo Center for Cosmology and Particle Physics New York University New York, NY 10003, USA albergo@nyu.edu Eric Vanden-Eijnden Courant Institute of Mathematical Sciences New York University New York, NY 10012, USA eve2@cims.nyu.edu |
| Pseudocode | No | The paper does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described in this paper. It mentions "lucidrains public diffusion code" in Appendix I, but this refers to a third-party implementation, not their own. |
| Open Datasets | Yes | ...image generation for CIFAR-10 and Image Net 32x32... CIFAR-10 (Krizhevsky et al., 2009) and Image Net 32 32 datasets (Deng et al., 2009; Van Den Oord et al., 2016)... Oxford flowers dataset (Nilsback & Zisserman, 2006)... A set of tabular datasets introduced by (Papamakarios et al., 2017)... |
| Dataset Splits | No | The paper states: "Results from the tabular experiments are displayed in Table 2, in which the negative log-likelihood averaged over a test set of held out data is computed." It mentions test sets but does not specify training, validation, or test split percentages or details needed for reproduction beyond stating "test data unseen during training". |
| Hardware Specification | Yes | We train an interpolant flow built from the UNet architecture from DDPM (Ho et al., 2020) on a single NVIDIA A100 GPU, which was previously impossible under maximum likelihood training of continuous time flows." and "All models were implemented on a single A100 GPU. |
| Software Dependencies | No | The paper mentions specific algorithms and implementations such as "Optimizer is performed on G(ˆv)", "numerical integration for sampling is done with the Dormand Prince", and "We built our models based off of the U-Net implementation provided by lucidrains public diffusion code, which we are grateful for https://github.com/lucidrains/denoising-diffusion-pytorch." However, it does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | The network for each model has 3 layers, each of width equal to 256 hidden units." and "Table 3: Hyperparameters and architecture for tabular datasets." and "Table 4: Hyperparameters and architecture for image datasets." These tables list detailed hyperparameters such as batch sizes, training steps, learning rates, network dimensions, and activation functions. |