PacGAN: The Power of Two Samples in Generative Adversarial Networks

Abstract

Generative adversarial networks (GANs) are innovative techniques for learning generative models of complex data distributions from samples. Despite remarkable recent improvements in generating realistic images, one of their major shortcomings is the fact that in practice, they tend to produce samples with little diversity, even when trained on diverse datasets. This phenomenon, known as mode collapse, has been the main focus of several recent advances in GANs. Yet there is little understanding of why mode collapse happens and why recently proposed approaches are able to mitigate mode collapse. We propose a principled approach to handling mode collapse, which we call packing. The main idea is to modify the discriminator to make decisions based on multiple samples from the same class, either real or artificially generated. We borrow analysis tools from binary hypothesis testing-in particular the seminal result of Blackwell [6]-to prove a fundamental connection between packing and mode collapse. We show that packing naturally penalizes generators with mode collapse, thereby favoring generator distributions with less mode collapse during the training process. Numerical experiments on benchmark datasets suggests that packing provides significant improvements in practice as well.

Normalized Diversification

“…In Fig. 4, we qualitatively demonstrate the results on synthetic point sets where our method achieve comparable or even better results over VEEGAN [39] and PacGAN [26] without using two or more real samples at a time.…”

Pixel-wise conditioned generative adversarial networks for image synthesis and completion

“…To avoid the problem, we exploit the recent PacGAN [16] technique: it consists in passing a set of samples to the discrimination function instead of a single one. PacGAN is intended to tackle the mode collapse problem in GAN training.…”

“…The underlying principle being that if a set of images are sampled from the same training set, they are very likely to be completely different, whereas if the generator experiences mode collapse, generated images are likely to be similar. In practice, we only give two samples to the discriminator, which is sufficient to overcome the loss of diversity as suggested in [16]. The resulting training procedure is summarized in Algorithm 2.…”

On-Line Adaptative Curriculum Learning for GANs

“…For example, CIFAR-10 has 50,000 training images and, assuming a batch size of 64, one epoch represents roughly 781 iterations for the generator. 4 σ(y) = 1/1 + e −y Modes (max 1000) KL DCGAN (Radford, Metz, and Chintala, 2015) 99.0 3.40 ALI (Dumoulin et al, 2016) 16.0 5.40 Unrolled GAN (Metz et al, 2016) 48.7 4.32 VEEGAN (Srivastava et al, 2017a) 150.0 2.95 PacGAN (Lin et al, 2017) 1000.0 ± 0.00 0.06 ± 1.0e −2 GAN+MINE (Belghazi et al, 2018) 1000.0 ± 0.00 0.05 ± 6.3e −3 acGAN (3D)…”

“…We report our results (averaged over 10 runs) in Table 2 and compare them with other existing baselines in the literature. Our method could recover all 1000 modes like PaCGAN (Lin et al, 2017) and MINE (Belghazi et al, 2018); these two approaches either increase the dimensionality of the generator inputs either by packing multiple samples or by adding a latent code vector which helps overcoming mode collapse. Generated samples are shown in Fig.…”