Day32
Generation
Dynamic Conditional Generation
Memory Network
Neural Turing Machine
Scheduled Sampling
Beam Search

Generation






Dynamic Conditional Generation








Memory Network
方式一



方式二




Neural Turing Machine
k对应产生attention









下图中，上是根据实际情况纠正，下是根据模型情况纠正。上面对未处理过问题时效果极差，下面模型预测效果差，一步错步步错。
调整其中参数权重将导致之前的训练失效


Scheduled Sampling



Beam Search
多条beam遍历，s

Generation Tasks
Improved GAN
related tasks
Improved GAN-Image Generation
GAN
Problems of GAN
Improved GAN
New paper from DeepMind
Semi-parametric Strategy
Image Generation
Message-Passing Strategy
Message Passing Strategy for Feature Sharing.
Domain Adaptation Strategy
Video generation
Text Effects Transfer
Scene graph Generation
Robot Control
Cross-Modal Zero-Shot Hashing retrieval
Reference
Acknowledgment


Reference[原文]: Joselynzhao.top & 夏木青 | Generation Tasks 

Improved GAN
如何从语言描述的场景来生成图像


related tasks

image synthesis





可以通过 RNN+GAN 生成实现
Improved GAN-Image Generation
GAN
Improved GAN-Image Generation

Image Generation from Scene Graphs (CVPR2018)
Stackgan: Text to photo-realistic image synthesis with

stacked generative adversarial networks. (ICCV 2017)

使用堆积的生成对抗网络实现从文本到真实图像的合成

Problems of GAN
carefully balance the capabilities between the generator and the discriminator

仔细平衡发生器和鉴别器之间的能力
unsuitable parameter settings can degrade the performance

不合适的参数设置会降低性能
Improved GAN
Evolutionary Generative Adversarial Networks (TEVC 2019)
New paper from DeepMind
Vector Quantized Variational AutoEncoder model（VQ-VAE2 wins bigGAN）

矢量量化变分自动编码器模型


Semi-parametric Strategy
半参数策略
Image Generation
Parametric models

end-to-end training of highly expressive models. （高度表达模型）
Nonparametric models （非参数模型）

extract materials from large real image datasets during testing

在测试的时候提取信息
Semi-parametric Strategy

Semi-parametric Image Synthesis (CVPR2018)
The problem of existing image generation methods （现有方法的问题）

give stunning results on limited domains such as

descriptions of birds, flowers or people.

Low-quality image of other domains.

作用领域有限
Using semi-parametric strategy to improve  Image generation method
Message-Passing Strategy
Message Passing Strategy for Feature Sharing.
用于特征共享的消息传递策略

Scene Graph Generation from Objects, Phrases and Region Captions(ICCV2017)

Domain Adaptation Strategy
区域适应策略

give stunning results on limited domains such as descriptions of birds, flowers or people.
Using Domain Adaptation to improve Image generation method  in more domains.
Video generation


输入 深度图像

输出 对手关节、姿势的估计


3D Hand Shape and Pose Estimation from a Single RGB Image（CVPR19 Oral）
Text Effects Transfer




创新的分布感知数据增强策略

TET-GAN: Text Effects Transfer via Stylization and Destylization

（AAAI2019）
Scene graph Generation


modeling these statistical correlations between object pairs and relationships can effectively regularize the semantic prediction space （structured knowledge graph ）

对这些对象和关系之间的统计相关性建模可以有效地规范语义预测空间（结构化知识图）
Knowledge-Embedded Routing Network for   Scene Graph Generation （CVPR2019）


Robot Control


Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection(2017)


Towards Accurate Task Accomplishment with Low-Cost Robotic Arms（CVPR2019）

Cross-Modal Zero-Shot Hashing retrieval
跨模态 zero-shot 哈希检索


Attribute-Guided Network for Cross-Modal Zero-Shot Hashing(T-NNLS19)
Reference

Image Generation from Scene Graphs (CVPR2018)
Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. (ICCV 2017)
Evolutionary Generative Adversarial Networks  (TEVC 2019)
Semi-parametric Image Synthesis (CVPR2018)
Scene Graph Generation from Objects, Phrases and Region Captions(ICCV2017)
3D Hand Shape and Pose Estimation from a Single RGB Image（CVPR19 Oral）
TET-GAN: Text Effects Transfer via Stylization and Destylization

（AAAI2019）
Knowledge-Embedded Routing Network for   Scene Graph Generation （CVPR2019）
Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection(2017)
Towards Accurate Task Accomplishment with Low-Cost Robotic Arms（CVPR2019）
Attribute-Guided Network for Cross-Modal Zero-Shot Hashing(T-NNLS19)

Acknowledgment
感谢本篇博文的内容提供者：崔亚文（Cui Yawen）· HPCL