【導(dǎo)讀】生成式對(duì)抗網(wǎng)絡(luò)模型(GAN)是基于深度學(xué)習(xí)的一種強(qiáng)大的生成模型,可以應(yīng)用于計(jì)算機(jī)視覺、自然語(yǔ)言處理、半監(jiān)督學(xué)習(xí)等重要領(lǐng)域。牛津大學(xué)Jakub Langr 撰寫了《GANs in Action》的著作,囊括了關(guān)于GAN的定義、訓(xùn)練、變體等,是學(xué)習(xí)生成式對(duì)抗網(wǎng)絡(luò)非常好的學(xué)習(xí)圖書,是關(guān)于GAN的最好的書籍之一。 網(wǎng)址: https://www./books/gans-in-action 【實(shí)戰(zhàn)GAN代碼和第一章便捷下載】 請(qǐng)關(guān)注專知公眾號(hào)(點(diǎn)擊上方藍(lán)色專知關(guān)注)
實(shí)戰(zhàn)GAN 深度學(xué)習(xí)系統(tǒng)在識(shí)別文本、圖像和視頻中的模式方面做得非常好。但事實(shí)證明,創(chuàng)建逼真的圖像、自然的句子和段落,或翻譯成地道英語(yǔ)的應(yīng)用程序是難以實(shí)現(xiàn)的。生成式對(duì)抗網(wǎng)絡(luò)(GANs)通過將兩個(gè)相互對(duì)抗的神經(jīng)網(wǎng)絡(luò)(一個(gè)生成內(nèi)容,另一個(gè)拒絕質(zhì)量較差的樣本)配對(duì),為這些挑戰(zhàn)提供了一個(gè)有希望的解決方案。 GAN GANs已經(jīng)取得了被認(rèn)為是人工系統(tǒng)不可能取得的顯著成果,例如能夠生成逼真的人臉、將涂鴉變成類似照片的圖像、將馬的視頻片段變成奔跑的斑馬。最重要的是,GANs學(xué)得很快,不需要費(fèi)力地標(biāo)注大量訓(xùn)練數(shù)據(jù)。 2014年由谷歌的Ian Goodfellow發(fā)明的生成式對(duì)抗網(wǎng)絡(luò)(GANs)是深度學(xué)習(xí)中最重要的創(chuàng)新之一。在GANs中,一個(gè)神經(jīng)網(wǎng)絡(luò)(生成器)生成內(nèi)容(圖像、句子等等),另一個(gè)神經(jīng)網(wǎng)絡(luò)(識(shí)別器)確定它們是否來(lái)自生成器,因此是“假的”,還是來(lái)自訓(xùn)練數(shù)據(jù)集,因此是“真實(shí)的”。在這兩個(gè)系統(tǒng)之間的相互作用中,當(dāng)生成器試圖欺騙鑒別器相信“贗品”是真實(shí)的時(shí),它會(huì)產(chǎn)生更真實(shí)的輸出。其結(jié)果是一個(gè)可以生成逼真圖像或自然文本和語(yǔ)音的生成器,以及一個(gè)訓(xùn)練有素的識(shí)別器,可以精確地識(shí)別和分類這類內(nèi)容。 本書簡(jiǎn)介 實(shí)戰(zhàn)GAN:用生成對(duì)抗式網(wǎng)絡(luò)進(jìn)行深度學(xué)習(xí)可以教會(huì)你如何構(gòu)建和訓(xùn)練自己的生成對(duì)抗式網(wǎng)絡(luò)。首先,您將了解生成模型以及GAN的工作原理,并概述它們的潛在用途。然后,隨著您探索GAN體系結(jié)構(gòu)的基礎(chǔ):生成器和鑒別器網(wǎng)絡(luò),您將開始構(gòu)建自己的簡(jiǎn)單的對(duì)抗系統(tǒng)。 通過本書迷人的示例和詳細(xì)的插圖,您將學(xué)習(xí)為不同的場(chǎng)景培訓(xùn)不同的GAN體系結(jié)構(gòu)。您將探索生成高分辨率圖像、圖像到圖像的轉(zhuǎn)換、對(duì)抗性學(xué)習(xí)以及目標(biāo)數(shù)據(jù)生成,從而使您的系統(tǒng)變得智能、有效和快速。 目錄: 1 INTRODUCTION TO GANS1.1 Introduction1.2 Prerequisites1.3 What Are Generative Adversarial Networks?1.3.1 GAN in Action1.3.2 GAN Training1.3.3 GAN Training Visualized1.3.4 Reaching Equilibrium1.3.5 The Pros and Cons of Studying GANs1.4 Applications of GANs1.5 Guide to this Book1.6 Summary2 AUTOENCODERS AS A PATH TO GANS 2.1 Why did we include this chapter?2.1.1 Generative learning is a new area for most2.1.2 Challenges of generative modelling2.1.3 An Important part of the literature today2.2 So what are autoencoders to GANs?2.3 What are the reasons behind autoencoders?2.4 Overview of Autoencoders2.5 Usage of autoencoders2.6 Unsupervised learning2.7 New take on an old idea2.8 Variational autoencoder (VAE)2.9 Code is life2.10 Summary3 YOUR FIRST GAN: GENERATING HANDWRITTEN DIGITS 3.1 Introduction3.1.1 Adversarial Training3.1.2 The Generator and the Discriminator3.1.3 GAN Training Algorithm3.2 Tutorial: Generating Handwritten Digits3.2.1 Import Statements3.2.2 The Generator3.2.3 The Discriminator3.2.4 Build the Model3.2.5 Training3.2.6 Outputting Sample Images3.2.7 Run the Model3.2.8 Inspecting the Results3.3 Conclusion3.4 Chapter Summary4 DEEP CONVOLUTIONAL GAN (DCGAN) 4.1 Introduction4.2 Convolutional Neural Networks (ConvNets)4.3 Brief History of the DCGAN4.4 Batch Normalization4.4.1 Computing Batch Normalization4.5 Tutorial: Generating Handwritten Digits with DCGAN4.5.1 Import Statements4.5.2 The Generator4.5.3 The Discriminator4.5.4 Build & Run the DCGAN4.5.5 Model Output4.6 Conclusion4.7 Chapter Summary5 TRAINING & COMMON CHALLENGES: GANING FOR SUCCESS 5.1 Evaluation5.1.1 Inception Score5.1.2 Fréchet Inception Distance5.2 Training challenges5.2.1 Network depth5.2.2 Game set-ups5.2.3 Min-Max GAN (MM-GAN)5.2.4 Non-Saturating GAN (NS-GAN)5.2.5 Summary of game setups 5.2.6 Training hacks5.3 Chapter summary6 PROGRESSING WITH GANS 6.1 Latent space interpolation 6.2 They grow up so fast6.2.1 Progressive Growing & Smoothing in of Higher Resolution Layers6.2.2 Minibatch Standard Deviation6.2.3 Equalized Learning Rate6.2.4 Pixel-wise Feature Normalization6.3 Summary of key innovations6.4 Tensorflow Hub and hands-on6.5 Practical Applications6.6 Chapter summary7 SEMI-SUPERVISED GAN 7.1 Introduction: “The GAN Zoo”7.2 Semi-Supervised GAN (SGAN)7.2.1 Why Semi-Supervised Learning?7.2.2 What is Semi-Supervised GAN?7.3 Tutorial: Implementing Semi-Supervised GAN7.3.1 Architecture Diagram7.3.2 Implementation7.3.3 Setup7.3.4 The Dataset7.3.5 The Generator7.3.6 The Discriminator7.3.7 Build the Model7.3.8 Training7.3.9 Train the Model7.3.10 Model Training and Test Accuracy7.3.11 Comparison to a Fully-Supervised Classifier7.4 Conclusion7.5 Summary8 CONDITIONAL GAN 8.1 Introduction8.2 Conditional GAN (CGAN)8.2.1 What is Conditional GAN?8.3 Tutorial: Implementing Conditional GAN8.3.1 Implementation8.3.2 Setup8.3.3 The Generator 8.3.4 Build the Model 8.3.5 Training 8.3.6 Outputting Sample Images 8.3.7 Train the Model 8.4 Inspecting the Output: Targeted Data Generation 8.5 Conclusion 8.6 Summary 9 CYCLEGAN 9.1 Introduction 9.2 Image-to-Image Translation 9.3 Cycle Consistent Loss: there and back aGAN 9.4 Adversarial Loss 9.5 Identity Loss9.6 Architecture9.7 CycleGAN architecture: building the network 9.8 Generator architecture9.9 Discriminator architecture9.10 Object Oriented Design of GANs9.11 Tutorial: CycleGAN9.12 Building the network9.13 Running CycleGAN9.14 Expansions, augmentations and applications9.15 Applications 9.16 Summary 10 ADVERSARIAL EXAMPLES 10.1 Introduction 10.2 Context of Adversarial Examples 10.3 Lies, Damned Lies and Distributions 10.4 Use and abuse of training 10.5 Signal and the noise 10.6 Not all hope is lost 10.7 Conclusion 10.8 Summary 11 PRACTICAL APPLICATIONS OF GANS 11.1 Introduction 11.2 GANs in Medicine 11.2.1 Using GANs to Improve Diagnostic Accuracy 11.3 GANs in Fashion 11.3.1 Using GANs to Design Fashion 11.4 Conclusion 11.5 Summary 12 LOOKING AHEAD 12.1 Introduction 12.2 Ethics 12.3 GAN Innovations 12.4 Relativistic GAN (RGAN) 12.4.1 Application 12.5 Self-Attention GAN (SAGAN) 12.5.1 Application 12.6 BigGAN 12.6.1 Application 12.7 Further reading 12.8 Looking Back & Closing Thoughts 12.9 Conclusion 12.10 Summary APPENDIXES APPENDIX A: TECHNICAL/DEPLOYMENTS作者介紹 Jakub Langr畢業(yè)于牛津大學(xué),并在牛津大學(xué)的計(jì)算服務(wù)院任教。他從2013年開始從事數(shù)據(jù)科學(xué)工作,最近開始擔(dān)任Filtered.com的數(shù)據(jù)科學(xué)技術(shù)主管和Mudano的研發(fā)數(shù)據(jù)科學(xué)家。Jakub是曼寧出版社出版的《GANs in Action》的合著作者,這本書是市場(chǎng)上目前唯一一本關(guān)于生成對(duì)抗網(wǎng)絡(luò)(GAN)的書。Jakub還在伯明翰大學(xué)參與設(shè)計(jì)并教授數(shù)據(jù)科學(xué)課程。 弗拉基米爾·博克(Vladimir Bok)是Intent Media的高級(jí)產(chǎn)品經(jīng)理。Intent Media是一家為領(lǐng)先旅游網(wǎng)站提供數(shù)據(jù)科學(xué)服務(wù)的公司,他幫助監(jiān)管該公司的機(jī)器學(xué)習(xí)研究和基礎(chǔ)設(shè)施團(tuán)隊(duì)。在此之前,他是微軟的項(xiàng)目經(jīng)理。弗拉迪米爾以優(yōu)異成績(jī)畢業(yè)于哈佛大學(xué)計(jì)算機(jī)科學(xué)專業(yè)。他曾在早期的金融科技公司擔(dān)任軟件工程師,包括PayPal聯(lián)合創(chuàng)始人Max Levchin創(chuàng)辦的一家公司,以及Y Combinator初創(chuàng)公司的數(shù)據(jù)科學(xué)家。 |
|
來(lái)自: LibraryPKU > 《機(jī)器學(xué)習(xí)》