Bridging the Gap Between f-GANs and Wasserstein GANs
Authors: Jiaming Song, Stefano Ermon
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate KL-WGAN empirically on distribution modeling, density estimation and image generation tasks. Empirical results demonstrate that KL-WGAN enjoys superior quantitative performance compared to its WGAN counterparts on several benchmarks. |
| Researcher Affiliation | Academia | Jiaming Song 1 Stefano Ermon 1 1Stanford University. Correspondence to: Jimaing Song <tsong@cs.stanford.edu>. |
| Pseudocode | Yes | Algorithm 1 Pseudo-code for KL-Wasserstein GAN |
| Open Source Code | Yes | We release code for our experiments (implemented in Py Torch) in https://github.com/ermongroup/f-wgan. To promote reproducible research, we include code in the supplementary material. |
| Open Datasets | Yes | We first demonstrate the effectiveness of KL-WGAN on synthetic and UCI benchmark datasets (Asuncion & Newman, 2007) considered in (Wenliang et al., 2018). The 2-d synthetic datasets include Mixture of Gaussians (Mo G), Banana, Ring, Square, Cosine and Funnel;... We use Red Wine, White Wine and Parkinsons from the UCI datasets. We further evaluate our KL-WGAN’s practical on image generation tasks on CIFAR10 and Celeb A datasets. |
| Dataset Splits | Yes | After training, we draw 5,000 samples from the generator and then evaluate two metrics over a fixed validation set. |
| Hardware Specification | No | The paper describes network architectures (Appendix D) and software (PyTorch), but does not specify the hardware (e.g., GPU models, CPU, or memory) used for running the experiments. |
| Software Dependencies | Yes | We release code for our experiments (implemented in Py Torch) in https://github.com/ermongroup/f-wgan. Our experiments are based on the Big GAN (Brock et al., 2018) Py Torch implementation. We present an implementation of KL-WGAN losses (in Py Torch) in Appendix B. |
| Experiment Setup | Yes | We use the same SNGAN (Miyato et al., 2018) arhictetures for WGAN and KL-WGANs, which uses spectral normalization to enforce Lipschitzness (detailed in Appendix D). For the Mo G, Square and Cosine datasets, we further show the estimated divergences over a batch of 256 samples in Figure 4. Appendix D. Network Architecture: For the synthetic experiments, we use the SNGAN (Miyato et al., 2018) architectures used in (Wenliang et al., 2018). ... For the image generation experiments, we use the Big GAN architectures and a fixed prior (the truncated normal distribution). |