人工智能培训

搜索

人工智能论文:MSG-GAN:多尺度梯度GAN,用于更稳定和同步的多尺度图像合成(MSG-GAN: Multi-Scale Gradients GAN for

[复制链接]
mhtq 发表于 2019-3-15 12:00:19 | 显示全部楼层 |阅读模式
mhtq 2019-3-15 12:00:19 923 0 显示全部楼层
人工智能论文:MSG-GAN:多尺度梯度GAN,用于更稳定和同步的多尺度图像合成(MSG-GAN: Multi-Scale Gradients GAN for more stable and synchronized  multi-scale image synthesis)虽然生成对抗网络(GAN)已经在图像合成任务中取得了巨大的成功,但它们的使用却非常难以使用,部分原因在于培训期间的不稳定性。这种不稳定性的一个普遍接受的原因是,由于训练期间的学习不平衡,从鉴别器到发生器的梯度很快就会变得无法提供信息。在这项工作中,我们提出了多尺度梯度生成对抗网络(MSG-GAN),这是一种简单但有效的技术,用于解决这个问题,从多个尺度的鉴别器到发生器的梯度流。该技术为生成同步的多尺度图像提供了稳定的方法。我们提出了一个非常直观的数学MSG-GAN框架实现,它使用了鉴别器计算中的连接操作。我们通过对CIFAR10和Oxford102花数据集的实验验证了我们的MSG-GAN方法的效果,并将其与进行多尺度图像合成的其他相关技术进行了比较。此外,我们还提供了CelebA-HQ数据集上的实验细节,用于合成1024 x 1024高分辨率图像。
While Generative Adversarial Networks (GANs) have seen huge successes inimage synthesis tasks, they are notoriously difficult to use, in part due toinstability during training.One commonly accepted reason for this instabilityis that gradients passing from the discriminator to the generator can quicklybecome uninformative, due to a learning imbalance during training.In thiswork, we propose the Multi-Scale Gradient Generative Adversarial Network(MSG-GAN), a simple but effective technique for addressing this problem whichallows the flow of gradients from the discriminator to the generator atmultiple scales.This technique provides a stable approach for generatingsynchronized multi-scale images.We present a very intuitive implementation ofthe mathematical MSG-GAN framework which uses the concatenation operation inthe discriminator computations.We empirically validate the effect of ourMSG-GAN approach through experiments on the CIFAR10 and Oxford102 flowersdatasets and compare it with other relevant techniques which performmulti-scale image synthesis.In addition, we also provide details of ourexperiment on CelebA-HQ dataset for synthesizing 1024 x 1024 high resolutionimages.人工智能论文:MSG-GAN:多尺度梯度GAN,用于更稳定和同步的多尺度图像合成(MSG-GAN: Multi-Scale Gradients GAN for more stable and synchronized  multi-scale image synthesis) oReR4ekg52LcGcf4.jpg
URL地址:https://arxiv.org/abs/1903.06048     ----pdf下载地址:https://arxiv.org/pdf/1903.06048    ----人工智能论文:MSG-GAN:多尺度梯度GAN,用于更稳定和同步的多尺度图像合成(MSG-GAN: Multi-Scale Gradients GAN for more stable and synchronized  multi-scale image synthesis)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

mhtq当前离线
新手上路

查看:923 | 回复:0

快速回复 返回顶部 返回列表