Hidden Convexity of Wasserstein GANs: Interpretable Generative Models with Closed-Form Solutions
Authors: Arda Sahiner, Tolga Ergen, Batu Ozturkler, Burak Bartan, John M. Pauly, Morteza Mardani, Mert Pilanci
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our observations are verified with numerical results demonstrating the power of the convex interpretation, with applications in progressive training of convex architectures corresponding to linear generators and quadratic-activation discriminators for Celeb A image generation. The code for our experiments is available at https://github.com/ardasahiner/Pro Co GAN. |
| Researcher Affiliation | Academia | Department of Electrical Engineering Stanford University Stanford, CA 94305, USA {sahiner,ergen,ozt,bbartan,pauly,morteza,pilanci}@stanford.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. It provides mathematical derivations and explanations of the method. |
| Open Source Code | Yes | The code for our experiments is available at https://github.com/ardasahiner/Pro Co GAN. For the sake of reproducibility, we provide all the experimental details (including preprocessing, hyperparameter optimization, extensive ablation studies, hardware requirements, and all other implementation details) in Appendix B as well as the source (https://github.com/ardasahiner/Pro Co GAN) to reproduce the experiments in the paper. |
| Open Datasets | Yes | We consider the task of generating images from the Celeb Faces Attributes Dataset (Celeb A) (Liu et al., 2015). |
| Dataset Splits | No | The paper does not provide specific dataset split information (exact percentages, sample counts, or detailed methodology) for training, validation, or testing splits beyond using the 50000 ground-truth images for training and comparison in FID calculation. |
| Hardware Specification | Yes | Both methods are trained with Pytorch (Paszke et al., 2019), where Pro Co GAN is trained with a single 12 GB NVIDIA Titan Xp GPU, while progressive GDA is trained with two of them. |
| Software Dependencies | No | The paper states, "Both methods are trained with Pytorch (Paszke et al., 2019)", but it does not specify the version number of Pytorch or any other software dependencies with their versions. |
| Experiment Setup | Yes | For the Progressive GDA baseline, we train the networks using Adam (Kingma & Ba, 2014), with α = 1e 3, β1 = 0, β2 = 0.99 and ϵ = 10 8, as is done in (Karras et al., 2017). [...] we use WGAN-GP loss with parameter λ = 10 and an additional penalty ϵdrift Ex px[D(x)2], where ϵdrift = 10 3. [...] m(i) d =(192, 192, 768, 3092, 3092) neurons at each stage, with fixed minibatches of size 16 for 15000 iterations per stage. [...] For Pro Co GAN, at each stage i, we use a fixed penalty β(i) d for the discriminator [...]. We keep (β(1) d , β(2) d , β(3) d ) fixed, and visualize the result of training two different sets of values of (β(4) d , β(5) d ) for Pro Co GAN. |