Difference-Seeking Generative Adversarial Network--Unseen Sample Generation

Authors: Yi Lin Sung, Sung-Hsien Hsieh, Soo-Chang Pei, Chun-Shien Lu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments are divided into three parts. The first one examines how the hyperparameter, α, influences the learned generator distribution, pg. In the second and third experiments, we obtain empirical results about semi-supervised learning and novelty detection, which are presented in Sec. 5.2 and Sec. 5.3, respectively. Note that the training procedure of the DSGAN can be improved by other extensions of GANs such as WGAN Arjovsky et al. (2017), WGAN-GP Gulrajani et al. (2017), EBGAN Zhao et al. (2017), and LSGAN Mao et al. (2017). In our method, the WGAN-GP was adopted for the stability of the DSGAN in training and reduction in the mode collapse.
Researcher Affiliation Academia Yi-Lin Sung Graduate Institute of Communication Engineering National Taiwan University, Taiwan, ROC Institute of Information Science, Academia Sinica r06942076@ntu.edu.tw Sung-Hsien Hsieh Institute of Information Science and Research Center for Information Technology Innovation, Academia Sinica, Taiwan, ROC parvaty316@hotmail.com Soo-Chang Pei Graduate Institute of Communication Engineering National Taiwan University, Taiwan, ROC peisc@ntu.edu.tw Chun-Shien Lu Institute of Information Science and Research Center for Information Technology Innovation, Academia Sinica, Taiwan, ROC lcs@iis.sinica.edu.tw
Pseudocode Yes The training procedure is illustrated in Algorithm 1 in Appendix A. [...] Algorithm 1 The training procedure of DSGAN using minibatch stochastic gradient descent.
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a code repository for the described methodology.
Open Datasets Yes Following the previous works, we apply the proposed DSGAN to semi-supervised learning on three benchmark datasets: MNIST Le Cun et al. (1998), SVHN Netzer et al. (2011), and CIFAR-10 Krizhevsky (2009).
Dataset Splits No The paper states: 'For evaluating the semi-supervised learning task, we used 60000/ 73257/ 50000 samples and 10000/ 26032/ 10000 samples from the MNIST/ SVHN/ CIFAR-10 datasets for the training and testing, respectively.' While it mentions training and testing sets, it does not explicitly define a separate validation set split or how it was used for hyperparameter tuning (e.g., exact percentages or sample counts for a validation set).
Hardware Specification Yes Table 2: Training times of our method and bad GAN. We only report the training time on MNIST, on which the authors of bad GAN applied Pixel CNN++. The experiments run on a NVIDIA 1080 Ti.
Software Dependencies No The paper mentions software like 'WGAN-GP' as a backbone and indicates implementations for 'MNIST, SVHN, and CIFAR-10' which would imply frameworks like PyTorch or TensorFlow, but it does not specify any version numbers for these software components, which is required for reproducibility.
Experiment Setup Yes The selected hyperparameters are listed in Table 5 in Appendix D.1. [...] Table 5: Hyperparameters in semi-supervised learning. Hyperparameters MNIST SVHN CIFAR-10 α 0.8 0.8 0.5 β 0.3 0.1 0.1 [...] We used k = 1 and α = 0.8 in experiments. [...] In the experiment, we first trained the VAE for 500 epochs and then we trained DSGAN for 500 epochs with m = 1.5 and w = 0.5. Third, we fixed the encoder and tuned the decoder with both positive and negative samples (generated by DSGAN) for 600 epochs.