The relativistic discriminator: a key element missing from standard GAN

Authors: Alexia Jolicoeur-Martineau

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we observe that 1) RGANs and Ra GANs are significantly more stable and generate higher quality data samples than their non-relativistic counterparts, 2) Standard Ra GAN with gradient penalty generate data of better quality than WGAN-GP while only requiring a single discriminator update per generator update (reducing the time taken for reaching the state-of-the-art by 400%), and 3) Ra GANs are able to generate plausible high resolutions images (256x256) from a very small sample (N=2011), while GAN and LSGAN cannot; these images are of significantly better quality than the ones generated by WGAN-GP and SGAN with spectral normalization. The code is freely available on https://github.com/Alexia JM/Relativistic GAN.
Researcher Affiliation Academia Alexia Jolicoeur-Martineau Lady Davis Institute MILA, Université de Montréal Montreal, Canada alexia.jolicoeur-martineau@mail.mcgill.ca
Pseudocode Yes H ALGORITHMS Algorithm 1 Training algorithm for non-saturating RGANs with symmetric loss functions
Open Source Code Yes The code is freely available on https://github.com/Alexia JM/Relativistic GAN.
Open Datasets Yes Experiments were conducted on the CIFAR-10 dataset (Krizhevsky, 2009) and the CAT dataset (Zhang et al., 2008).
Dataset Splits No The paper does not explicitly provide training/test/validation dataset splits needed to reproduce the experiment. It mentions using CIFAR-10 and CAT datasets but does not detail how they were split for training, validation, and testing.
Hardware Specification No The paper does not provide SPECIFIC HARDWARE DETAILS (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No Code was written in Pytorch (Paszke et al., 2017) and models were trained using the Adam optimizer (Kingma & Ba, 2014). The paper mentions PyTorch but does not specify its version number or any other software dependencies with specific versions.
Experiment Setup Yes We used the following two known stable setups: (DCGAN setup) lr = .0002, n D = 1, β1 = .50 and β2 = .999 (Radford et al., 2015), and (WGAN-GP setup) lr = .0001, n D = 5, β1 = .50 and β2 = .9 (Gulrajani et al., 2017), where lr is the learning rate, n D is the number of discriminator updates per generator update, and β1, β2 are the ADAM momentum parameters. For optimal stability, we used batch norm (Ioffe & Szegedy, 2015) in G and spectral norm (Miyato et al., 2018) in D.