人工智能培训

搜索

人工智能论文:模式寻求生成的对抗网络的多样化图像合成(Mode Seeking Generative Adversarial Networks for Dive

[复制链接]
xibao 发表于 2019-3-15 13:26:50 | 显示全部楼层 |阅读模式
xibao 2019-3-15 13:26:50 525 0 显示全部楼层
人工智能论文:模式寻求生成的对抗网络的多样化图像合成(Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis)大多数条件生成任务在单一的上下文环境下期望产生不同的输出。然而,条件生成对抗网络(cGAN)通常关注先前的条件信息并忽略输入噪声向量,这有助于输出变化。最近解决cGAN模式崩溃问题的尝试通常是特定任务和计算上昂贵的。在这项工作中,我们提出了一个简单而有效的管理术语来解决cGAN的模式崩溃问题。所提出的方法明确地最大化所生成的图像之间的距离与相应的潜在代码的比率,从而鼓励生成器在训练期间探索更多的小模式。寻求正则化项的这种模式很容易适用于各种条件生成任务,其中包括培训开销或修改原始网络结构。我们在三个条件图像合成任务上验证所提出的算法,包括分类生成,图像到图像转换以及具有不同基线模型的文本到图像合成。定性和定量结果均证明了所提出的正则化方法在不损失质量的情况下改善多样性的有效性。
Most conditional generation tasks expect diverse outputs given a singleconditional context.However, conditional generative adversarial networks(cGANs) often focus on the prior conditional information and ignore the inputnoise vectors, which contribute to the output variations.Recent attempts toresolve the mode collapse issue for cGANs are usually task-specific andcomputationally expensive.In this work, we propose a simple yet effectiveregularization term to address the mode collapse issue for cGANs.The proposedmethod explicitly maximizes the ratio of the distance between generated imageswith respect to the corresponding latent codes, thus encouraging the generatorsto explore more minor modes during training.This mode seeking regularizationterm is readily applicable to various conditional generation tasks withoutimposing training overhead or modifying the original network structures.Wevalidate the proposed algorithm on three conditional image synthesis tasksincluding categorical generation, image-to-image translation, and text-to-imagesynthesis with different baseline models.Both qualitative and quantitativeresults demonstrate the effectiveness of the proposed regularization method forimproving diversity without loss of quality.人工智能论文:模式寻求生成的对抗网络的多样化图像合成(Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis) A001V48S1S0H9Ss1.jpg
URL地址:https://arxiv.org/abs/1903.05628     ----pdf下载地址:https://arxiv.org/pdf/1903.05628    ----人工智能论文:模式寻求生成的对抗网络的多样化图像合成(Mode Seeking Generative Adversarial Networks for Diverse 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

xibao当前离线
新手上路

查看:525 | 回复:0

快速回复 返回顶部 返回列表