Improving GAN Training with Probability Ratio Clipping and Sample Reweighting

Authors: Yue Wu, Pan Zhou, Andrew G. Wilson, Eric Xing, Zhiting Hu

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on three unsupervised generation tasks, including image generation, text generation, and text style transfer. The three tasks apply GANs to model different data modalities, namely, image, text, and neural hidden representations, respectively. Our approach consistently offers improvement over the state-of-the-arts on all tasks. See appendix for all experimental details.
Researcher Affiliation Collaboration Yue Wu1, Pan Zhou2, Andrew Gordon Wilson3, Eric P. Xing1,4, Zhiting Hu1,5 1Carnegie Mellon University, 2National University of Singapore, 3New York University 4Petuum Inc., 5UC San Diego
Pseudocode Yes Algorithm 1 GAN Training with Probability Ratio Clipping and Sampling Re-weighting
Open Source Code Yes Code available at: github.com/Holmeswww/PPOGAN
Open Datasets Yes We first use the popular CIFAR-10 benchmark for evaluation... CIFAR-10 [28] contains 50K images of sizes 32 × 32. ... We then evaluate our method on the EMNLP2017 WMT News, a large real text data used for text GAN studies [16, 35]. ... We use the standard Yelp review dataset, and the ground truth output text provided by [30].
Dataset Splits No The dataset consists of 270K/10K training/test sentences with a maximum length of 51 and a vocabulary size of 5,255.
Hardware Specification No The paper does not provide any specific details about the hardware specifications (e.g., GPU models, CPU types, memory) used to conduct the experiments.
Software Dependencies No The paper mentions using specific algorithms and architectures such as 'Gumbel-softmax approximation [25, 31]' and implementing based on 'Rel GAN [35] architecture', but it does not specify software versions for any frameworks (e.g., PyTorch, TensorFlow) or libraries used.
Experiment Setup Yes For each iteration, we update both generator and discriminator for 5 times. ... Our full approach and clipping only use an update ratio of 5:5, because the probability ratio clipping that discourages large generator updates allows us to update the generator more frequently, which is desirable. ... the generator is initialized with maximum likelihood (MLE) pre-training.