Tempered Adversarial Networks
Authors: Mehdi S. M. Sajjadi, Giambattista Parascandolo, Arash Mehrjou, Bernhard Schölkopf
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In a number of experiments, we show that this can improve quality, stability and/or convergence speed across a range of different GAN architectures (DCGAN, LSGAN, WGAN-GP). |
| Researcher Affiliation | Academia | 1Max Planck Institute for Intelligent Systems, T ubingen, Germany 2Max Planck ETH Center for Learning Systems, Z urich, Switzerland. |
| Pseudocode | No | The paper describes the model architecture and training process in text and figures, but it does not include a dedicated pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain an explicit statement or a link indicating that the source code for their proposed method is publicly available. |
| Open Datasets | Yes | We begin with the original GAN variant on the classical MNIST dataset. Since MNIST only has 10 main modes, it is not an adequate test for the mode collapse problem in GANs. To alleviate this, a color MNIST variant has been proposed (Srivastava et al., 2017). On the Celeb A dataset (Liu et al., 2015)... our experiment on the Cifar-10 dataset shows that the same lens with the same hyperparameters also works well with WGAN-GP. |
| Dataset Splits | No | The paper mentions datasets used for experiments but does not explicitly provide details about training, validation, and test splits (e.g., percentages or sample counts for each split). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments (e.g., GPU models, CPU types, or memory specifications). |
| Software Dependencies | No | The paper mentions optimizers used (Adam, RMSProp) but does not provide specific version numbers for any software dependencies, libraries, or frameworks used in the implementation. |
| Experiment Setup | Yes | We train using batch sizes of 32 and 64, a learning rate of 10 4 and we initialize the networks with the Xavier initialization (Gloriot & Bengio, 2010). For the experiments with the original GAN loss, we use the DCGAN architecture along with its common tweaks (Radford et al., 2016), namely, strided convolutions instead of pooling layers, applying batch normalization in both networks, using Re LU in the generator and leaky Re LU in the discriminator, and Adam (Kingma & Ba, 2015) as the optimizer. In all experiments, we set K = 105 unless specified otherwise... The image size in all experiments is 32 32 pixels with 1 color channel for MNIST and 3 color channels for all other experiments. |