Multi-Agent Diverse Generative Adversarial Networks

Ghosh, Arnab; Kulharia, Viveka; Namboodiri, Vinay P.; Torr, Philip H. S.; Dokania, Puneet K.

doi:10.48550/arxiv.1704.02906

Cited by 20 publications

(38 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To illustrate the functioning of DoPaNet, we first set up two low-dimensional experiments (Section 4.1) using Gaussian Mixture Models (GMMs) as the target probability density function: 1D GMM and 2D GMM. For the 1D Gaussian Mixture case, we compare DoPaNet's robustness against other approaches by reproducing the experiment setting detailed in (Ghosh et al, 2017) and we outperform all competing methods both qualitatively and quantitatively. We also show DoPaNet's performance using multiple discriminators and show how the training dynamics change according to the number of discriminators.…”

Section: Methodsmentioning

confidence: 99%

“…Several works propose using multiple generators (Arora et al, 2017a;Ghosh et al, 2016;Liu & Tuzel, 2016). For instance, MAD-GAN (Ghosh et al, 2017) improves the learning by compelling the generators to produce diverse modes implicitly using the discriminator. This is achieved by requiring the discriminator to identify the generator that produced the fake samples along with recognizing fake samples from reals.…”

Section: Related Workmentioning

confidence: 99%

“…In DoPaNet, the role of the classifier Q is to partition both the real and generated data-points into different clusters or modes, and each discriminator is consequently only trained on a separate cluster. In order to fully understand how this helps the training, we experimented with two toy datasets obtained using mixture of Gaussian variants: a 1D GMM with 5 modes, as used in (Ghosh et al, 2017), and a 2D circular GMM with 3 and 8 modes on the unit circle. First, we reproduced the 1D setting in (Ghosh et al, 2017) with 5 modes at [10,20,60,80,110] and standard deviations [3, 3, 2, 2, 1] respectively and we compare to the numbers reported in that paper in Table 1.…”

Section: Synthetic Low Dimensional Distributionsmentioning

confidence: 99%

“…In order to fully understand how this helps the training, we experimented with two toy datasets obtained using mixture of Gaussian variants: a 1D GMM with 5 modes, as used in (Ghosh et al, 2017), and a 2D circular GMM with 3 and 8 modes on the unit circle. First, we reproduced the 1D setting in (Ghosh et al, 2017) with 5 modes at [10,20,60,80,110] and standard deviations [3, 3, 2, 2, 1] respectively and we compare to the numbers reported in that paper in Table 1. We sampled 65, 536 data points each from the real distribution and the generator distribution.…”

Section: Synthetic Low Dimensional Distributionsmentioning

confidence: 99%

“…However, the generator might miss some modes of the distribution even after reaching the equilibrium as it can simply fool the discriminator by generating from only few modes of the real distribution (Goodfellow, 2016;Che et al, 2017;Chen et al, 2016;Salimans et al, 2016), and hence producing a limited diversity in samples. To address this problem, the literature explores two main approaches: Improving GAN learning to practically reach a better optimum Metz et al, 2017;Salimans et al, 2016;Gulrajani et al, 2017;Berthelot et al, 2017), or explicitly forcing GANs to produce various modes by design (Chen et al, 2016;Ghosh et al, 2017;Durugkar et al, 2017;Che et al, 2017;Liu & Tuzel, 2016). We hereby follow the latter strategy and propose a new way of dealing with GAN mode collapse.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Domain Partitioning Network

Csaba¹,

Boukhayma²,

Kulharia³

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

Standard adversarial training involves two agents, namely a generator and a discriminator, playing a mini-max game. However, even if the players converge to an equilibrium, the generator may only recover a part of the target data distribution, in a situation commonly referred to as mode collapse.In this work, we present the Domain Partitioning Network (DoPaNet), a new approach to deal with mode collapse in generative adversarial learning. We employ multiple discriminators, each encouraging the generator to cover a different part of the target distribution. To ensure these parts do not overlap and collapse into the same mode, we add a classifier as a third agent in the game. The classifier decides which discriminator the generator is trained against for each sample. Through experiments on toy examples and real images, we show the merits of DoPaNet in covering the real distribution and its superiority with respect to the competing methods. Besides, we also show that we can control the modes from which samples are generated using DoPaNet.

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Synthetic Low Dimensional Distributionsmentioning

confidence: 99%

Section: Synthetic Low Dimensional Distributionsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Domain Partitioning Network

Csaba¹,

Boukhayma²,

Kulharia³

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

Multimodal Unsupervised Image-to-Image Translation

Huang

Belongie

et al. 2018

Lecture Notes in Computer Science

2,185

2,449

View full text Add to dashboard Cite

Unsupervised image-to-image translation is an important and challenging problem in computer vision. Given an image in the source domain, the goal is to learn the conditional distribution of corresponding images in the target domain, without seeing any examples of corresponding image pairs. While this conditional distribution is inherently multimodal, existing approaches make an overly simplified assumption, modeling it as a deterministic one-to-one mapping. As a result, they fail to generate diverse outputs from a given source domain image. To address this limitation, we propose a Multimodal Unsupervised Image-to-image Translation (MUNIT) framework. We assume that the image representation can be decomposed into a content code that is domain-invariant, and a style code that captures domain-specific properties. To translate an image to another domain, we recombine its content code with a random style code sampled from the style space of the target domain. We analyze the proposed framework and establish several theoretical results. Extensive experiments with comparisons to state-of-the-art approaches further demonstrate the advantage of the proposed framework. Moreover, our framework allows users to control the style of translation outputs by providing an example style image. Code and pretrained models are available at https://github.com/nvlabs/MUNIT.

show abstract

Fictitious GAN: Training GANs with Historical Models

Yin

Chen

et al. 2018

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Generative adversarial networks (GANs) are powerful tools for learning generative models. In practice, the training may suffer from lack of convergence. GANs are commonly viewed as a two-player zerosum game between two neural networks. Here, we leverage this game theoretic view to study the convergence behavior of the training process. Inspired by the fictitious play learning process, a novel training method, referred to as Fictitious GAN, is introduced. Fictitious GAN trains the deep neural networks using a mixture of historical models. Specifically, the discriminator (resp. generator) is updated according to the bestresponse to the mixture outputs from a sequence of previously trained generators (resp. discriminators). It is shown that Fictitious GAN can effectively resolve some convergence issues that cannot be resolved by the standard training approach. It is proved that asymptotically the average of the generator outputs has the same distribution as the data samples.The global optimum point is reached when the generated distribution p g , which is the distribution of G(z) given z ∼ p z (z), is equal to the data distribution. The optimal point is reached based on the assumption that the discriminator and ⋆ First three authors have equal contributions.

show abstract

Multi-Agent Diverse Generative Adversarial Networks

Cited by 20 publications

References 0 publications

Domain Partitioning Network

Domain Partitioning Network

Multimodal Unsupervised Image-to-Image Translation

Fictitious GAN: Training GANs with Historical Models

Contact Info

Product

Resources

About