AmbientGAN: Generative models from lossy measurements

Authors: Ashish Bora, Eric Price, Alexandros G. Dimakis

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On three benchmark datasets, and for various measurement models, we demonstrate substantial qualitative and quantitative improvements. Generative models trained with our method can obtain 2-4x higher inception scores than the baselines.
Researcher Affiliation Academia Ashish Bora Department of Computer Science University of Texas at Austin ashish.bora@utexas.edu Eric Price Department of Computer Science University of Texas at Austin ecprice@cs.utexas.edu Alexandros G. Dimakis Department of Electrical and Computer Engineering University of Texas at Austin dimakis@austin.utexas.edu
Pseudocode No The paper describes algorithms in prose and mathematical formulations but does not include formal pseudocode blocks or algorithms.
Open Source Code Yes code reused from https://github.com/carpedm20/DCGAN-tensorflow; code reused from https://github.com/igul222/improved_wgan_training
Open Datasets Yes MNIST is a dataset of 28 28 images of handwritten digits [Le Cun et al. (1998)]. Celeb A is a dataset of face images of celebrities [Liu et al. (2015)]. We use an aligned and cropped version where each image is 64 64 RGB. The CIFAR-10 dataset consists of 32 32 RGB images from 10 different classes [Krizhevsky & Hinton (2009)].
Dataset Splits No The paper mentions training, testing, and standard datasets but does not explicitly provide details about training/validation/test splits, percentages, or sample counts for reproducibility beyond mentioning standard benchmarks.
Hardware Specification No The paper does not explicitly describe the hardware used for running the experiments. It only mentions general setups implicitly by discussing training of neural networks.
Software Dependencies No The paper mentions the use of TensorFlow (implied by github links such as tensorflow.org and carpedm20/DCGAN-tensorflow) and specific models/codebases (DCGAN, WGANGP, ACWGANGP) but does not provide specific version numbers for TensorFlow or any other software dependencies.
Experiment Setup Yes For the MNIST dataset, we use two GAN models. The first model is a conditional DCGAN which follows the architecture in [Radford et al. (2015)], while the second model is an unconditional Wasserstein GAN with gradient penalty (WGANGP) which follows the architecture in [Gulrajani et al. (2017)]. For the celeb A dataset, we use an unconditional DCGAN and follow the architecture in [Radford et al. (2015)]. For the CIFAR-10 dataset, we use an Auxiliary Classifier Wasserstein GAN with gradient penalty (ACWGANGP) which follows the residual architecture in [Gulrajani et al. (2017)]. More details on architectures and hyperparameters can be found in the appendix.