Cost Ensemble with Gradient Selecting for GANs
Authors: Minghui Liu, Jiali Deng, Meiyi Yang, Xuan Cheng, Nianbo Liu, Ming Liu, Xiaomin Wang
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide mathematical statements to prove our assumptions and conduct extensive experiments to verify the performance. The results show that CES-GAN is lightweight and more effective for fighting against the mode collapse problem than similar works. Meanwhile, we conduct extensive experiments on various well-known datasets to evaluate the contributions. Results show that CES-GAN achieves excellent performance boosts and is less prone to mode collapse. |
| Researcher Affiliation | Academia | University of Electronic Science and Technology of China minghuiliu@std.uestc.edu.cn, dengjiali@std.uestc.edu.cn, meiyiyang@std.uestc.edu.cn, cs xuancheng@std.uestc.edu.cn, liunb@uestc.edu.cn, csmliu@uestc.edu.cn, xmwang@uestc.edu.cn |
| Pseudocode | Yes | Algorithm 1 The gradient selecting mechanism for CES-GAN in each training iteration |
| Open Source Code | No | The paper does not provide any explicit statement or link indicating the availability of its source code. |
| Open Datasets | Yes | Experiments with two popular cost functions in GANs are conducted to explore the reason for that, including CTGAN [Wei et al., 2018] and BEGAN [Berthelot et al., 2017]. As shown in Figure 1(a), the Wasserstein distance between the real distribution and the generated distribution of the CTGAN is always smaller than the one with BEGAN in the whole training procedure using the CIFAR-10 training dataset [Krizhevsky and Hinton, 2009]. Instead, the distance is always larger than the one with BEGAN on the Celeb A dataset [Krizhevsky and Hinton, 2009]... The proposed framework is first evaluated on three small but well-known estimating the performance of combat the mode collapse problem datasets: the prevalent MNIST [Le Cun et al., 1998], the Stacked MNIST [Lin et al., 2018], and the synthetic dataset. We compare with the SOTA methods both quantitatively and qualitatively on the CIFAR-10 dataset. Besides, Results on a large-scale dataset named Celeb A [Liu et al., 2015; Karras et al., 2018] are described and compared with other works. Another large-scale dataset called Image Net will be shown in Appendix D. |
| Dataset Splits | Yes | The MNIST of handwritten digits is a subset of a more extensive set available from NIST, which provides 70,000 examples in total, and 10,000 of them are left out for the testing. For the semi-supervised learning approach, same with [Wei et al., 2018], we follow the standard training/test split of the dataset but use only 4,000 labels in training. |
| Hardware Specification | Yes | All the experiments are implemented and evaluated with Pytorch on 8 Nvidia Geforce GTX 1080 Ti [Paszke et al., 2019]. |
| Software Dependencies | No | The paper mentions 'Pytorch' as the implementation framework but does not specify its version number or any other software dependencies with version numbers. |
| Experiment Setup | Yes | Adam optimized with an initial learning rate of 1e-4 has been used to train them, and no transform or data augmentation has been utilized in these experiments. The prior noise input to the generator follows the random Gaussian distribution with 128 sizes. The number of samples used to train the networks is 128,000 samples, with a batch size of 64 samples. train GANs with 100,000 total samples and a batch size of 100 pieces. Regular data augmentation with flipping the images horizontally and randomly translating the images within -2 and 2 pixels is utilized. |