Stein Latent Optimization for Generative Adversarial Networks
Authors: Uiwon Hwang, Heeseung Kim, Dahuin Jung, Hyemi Jang, Hyungyu Lee, Sungroh Yoon
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We performed experiments on various real-world datasets including MNIST (Le Cun et al., 1998), Fashion-MNIST (Xiao et al., 2017), CIFAR-10 (Krizhevsky et al., 2009), Celeb A (Liu et al., 2015), Celeb A-HQ (Karras et al., 2017), and AFHQ (Choi et al., 2020) using architectures such as DCGAN (Radford et al., 2016), Res GAN (Gulrajani et al., 2017), and Style GAN2 (Karras et al., 2020). Through experiments, we verified that the proposed method outperforms existing unsupervised conditional GANs in unsupervised conditional generation on datasets with balanced or imbalanced attributes. |
| Researcher Affiliation | Academia | Uiwon Hwang1, Heeseung Kim1, Dahuin Jung1, Hyemi Jang1, Hyungyu Lee1, Sungroh Yoon1,2, 1Department of Electrical and Computer Engineering, 2AIIS, ASRI, INMC, ISRC, NSI, and Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul 08826, Korea {uiwon.hwang, gmltmd789, annajung0625, wkdal9512, rucy74, sryoon}@snu.ac.kr |
| Pseudocode | Yes | Algorithm 1 Training procedure of SLOGAN; Algorithm 2 Intra-cluster FID |
| Open Source Code | Yes | Code is available at https://github.com/shinyflight/SLOGAN |
| Open Datasets | Yes | We used the MNIST (Le Cun et al., 1998), Fashion-MNIST (Xiao et al., 2017), CIFAR-10 (Krizhevsky et al., 2009), Celeb A (Liu et al., 2015), Celeb A-HQ (Karras et al., 2017), and AFHQ (Choi et al., 2020) datasets to evaluate the proposed method. |
| Dataset Splits | No | The paper specifies training and test set sizes for some datasets (e.g., 60,000 training and 10,000 test for MNIST/FMNIST), but does not explicitly provide details for a validation split. |
| Hardware Specification | Yes | The experiments herein were conducted on a machine equipped with an Intel Xeon Gold 6242 CPU and an NVIDIA Quadro RTX 8000 GPU. |
| Software Dependencies | Yes | The code is implemented in Python 3.7 and Tensorflow 1.14 (Abadi et al., 2016). |
| Experiment Setup | Yes | We set the learning rate of G as η and the learning rate of Σ as γ. Throughout the experiments, we set the learning rate of E to η, and D to 4η using the two-timescale update rule (TTUR) (Heusel et al., 2017). We set the learning rate of µ to 10γ, and the learning rate of ρ to γ. We set B to 64 and the number of training steps to 100k. |