Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields
Authors: Thomas Unterthiner, Bernhard Nessler, Calvin Seward, Günter Klambauer, Martin Heusel, Hubert Ramsauer, Sepp Hochreiter
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show the efficacy of Coulomb GANs on LSUN bedrooms, Celeb A faces, CIFAR-10 and the Google Billion Word text generation. In all of our experiments, we used a low-dimensional Plummer Kernel of dimensionality d = 3. This kernel both gave best computational performance and has low risk of running into numerical issues. We used a batch size of 128. To evaluate the quality of a GAN, the FID metric as proposed by Heusel et al. (2017) was calculated by using 50k samples drawn from the generator, while the training set statistics were calculated using the whole training set. |
| Researcher Affiliation | Collaboration | Thomas Unterthiner1 Bernhard Nessler1 Calvin Seward1,2 Günter Klambauer1 Martin Heusel1 Hubert Ramsauer1 Sepp Hochreiter1 1LIT AI Lab & Institute of Bioinformatics, Johannes Kepler University Linz, Austria 2Zalando Research, Mühlenstraße 25, 10243 Berlin, Germany |
| Pseudocode | Yes | A.6 PSEUDOCODE FOR COULOMB GANS Algorithm 1 Minibatch stochastic gradient descent training of Coulomb GANs for updating the the discriminator weights w and the generator weights θ. |
| Open Source Code | Yes | The implementation used for these experiments is available online1. 1 www.github.com/bioinf-jku/coulomb_gan |
| Open Datasets | Yes | We show the efficacy of Coulomb GANs on LSUN bedrooms, Celeb A faces, CIFAR-10 and the Google Billion Word text generation. The cropped and centered images of celebrities from the Large-scale Celeb Faces Attributes ( Celeb A ) data set (Liu et al., 2015), the LSUN bedrooms data set consists of over 3 million 64x64 pixel images of the bedrooms category of the large scale image database LSUN (Yu et al., 2015) as well as the CIFAR-10 data set. |
| Dataset Splits | No | The paper mentions calculating 'training set statistics' and using the FID metric to stop learning, but does not explicitly provide specific train/validation/test dataset splits with percentages or counts for the input data. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, or memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and refers to frameworks like Tensorflow or Theano, but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | In all of our experiments, we used a low-dimensional Plummer Kernel of dimensionality d = 3. ... We used a batch size of 128. For the Plummer kernel, ϵ was set to 1. We used the Adam optimizer with a learning rate of 10 4 for the generator and 5 10 5 for the discriminator. ... For regularization we used an L2 weight decay term with a weighting factor of 10 7. |