Woodbury Transformations for Deep Generative Flows
Authors: You Lu, Bert Huang
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we compare the performance of Woodbury transformations against other modern flow architectures, measuring running time, bit per-dimension (log2-likelihood), and sample quality. We train with the CIFAR-10 [23] and Image Net [31] datasets. |
| Researcher Affiliation | Academia | You Lu Department of Computer Science Virginia Tech Blacksburg, VA you.lu@vt.edu Bert Huang Department of Computer Science Tufts University Medford, MA bert@cs.tufts.edu |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | The paper does not provide a statement about releasing source code or a direct link to a code repository for the described methodology. |
| Open Datasets | Yes | We train with the CIFAR-10 [23] and Image Net [31] datasets. We train Glow and Woodbury-Glow on the Celeb A-HQ dataset [19]. |
| Dataset Splits | No | The paper mentions training and evaluating on specific datasets (CIFAR-10, ImageNet, Celeb A-HQ) and reports 'test-set likelihoods' but does not explicitly provide the specific training, validation, and test dataset splits (e.g., percentages, sample counts, or clear citation to a predefined split methodology). |
| Hardware Specification | Yes | For fair comparison, we implement all methods in Pytorch and run them on an Nvidia Titan V GPU. |
| Software Dependencies | No | The paper mentions 'Pytorch' but does not provide specific version numbers for Pytorch or any other software dependencies. |
| Experiment Setup | Yes | For 32 32 images, we set the number of levels to L = 3 and the number of steps per-level to K = 8. For 64 64 images, we use L = 4 and K = 16. More details are in the appendix. For Woodbury transformations, we fix the latent dimension d = 16. We use 5bit images and set the size of images to be 64 64, 128 128, and 256 256. Detailed parameter settings are in the appendix. |