Memorization Precedes Generation: Learning Unsupervised GANs with Memory Networks

Authors: Youngjin Kim, Minjung Kim, Gunhee Kim

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

Reproducibility Variable Result LLM Response
Research Type Experimental We propose a novel end-to-end GAN model named memory GAN, which involves a memory network that is unsupervisedly trainable and integrable to many existing GAN models. With evaluations on multiple datasets such as Fashion-MNIST, Celeb A, CIFAR10, and Chairs, we show that our model is probabilistically interpretable, and generates realistic image samples of high visual fidelity. The memory GAN also achieves the state-of-the-art inception scores over unsupervised GAN models on the CIFAR10 dataset, without any optimization tricks and weaker divergences.
Researcher Affiliation Academia Youngjin Kim, Minjung Kim & Gunhee Kim Department of Computer Science and Engineering, Seoul National University, Seoul, Korea {youngjin.kim,minjung.kim,gunhee.kim}@vision.snu.ac.kr
Pseudocode Yes Algorithm 1 Training algorithm of memory GAN. φ is parameters of discriminator, θ is parameters of generator. α = 0.5, η = 2 10 4.
Open Source Code Yes The code is available at https: //github.com/whyjay/memory GAN.
Open Datasets Yes Fashion-MNIST (Xiao et al., 2017), Celeb A (Liu et al., 2015), CIFAR10 (Krizhevsky, 2009), and Chairs (Aubry et al., 2014).
Dataset Splits No The paper uses common datasets like CIFAR10, which have predefined splits, but it does not explicitly state the percentages or counts for training/validation/test splits within the text. It only mentions 'Test sample size' for inception scores.
Hardware Specification No The paper states training times and parameters but does not specify any hardware details like GPU/CPU models or memory.
Software Dependencies No The paper mentions using DCGAN, WGAN-GP ResNet, Adam optimizer, but does not provide specific version numbers for any software, libraries, or frameworks.
Experiment Setup Yes We use minibatches of size 64, a learning rate of 2 10 4, and Adam (Kingma & Ba, 2014) optimizer for all experiments. More implementation details can be found in Appendix B. For MNIST and Fashion-MNIST, we use DCGAN (Radford et al., 2016) for the inference network and generator, while for CIFAR10 and Celeb A, we use the WGAN-GP Res Net (Gulrajani et al., 2017) with minor changes such as using layer normalization (Ba et al., 2016) instead of batch normalization and using ELU activation functions instead of Re LU and Leaky Re LU.