book

Generative Deep Learning

by David Foster

June 2019

Intermediate to advanced

327 pages

7h 36m

English

O'Reilly Media, Inc.

Read now

Unlock full access

Objective and ApproachPrerequisitesOther ResourcesConventions Used in This BookUsing Code ExamplesO’Reilly Online LearningHow to Contact UsAcknowledgments
What Is Generative Modeling?Generative Versus Discriminative ModelingAdvances in Machine LearningThe Rise of Generative ModelingThe Generative Modeling FrameworkProbabilistic Generative ModelsHello Wrodl!Your First Probabilistic Generative ModelNaive BayesHello Wrodl! ContinuedThe Challenges of Generative ModelingRepresentation LearningSetting Up Your EnvironmentSummary
Structured and Unstructured DataDeep Neural NetworksKeras and TensorFlowYour First Deep Neural NetworkLoading the DataBuilding the ModelCompiling the ModelTraining the ModelEvaluating the ModelImproving the ModelConvolutional LayersBatch NormalizationDropout LayersPutting It All TogetherSummary
The Art ExhibitionAutoencodersYour First AutoencoderThe EncoderThe DecoderJoining the Encoder to the DecoderAnalysis of the AutoencoderThe Variational Art ExhibitionBuilding a Variational AutoencoderThe EncoderThe Loss FunctionAnalysis of the Variational AutoencoderUsing VAEs to Generate FacesTraining the VAEAnalysis of the VAEGenerating New FacesLatent Space ArithmeticMorphing Between FacesSummary
GanimalsIntroduction to GANsYour First GANThe DiscriminatorThe GeneratorTraining the GANGAN ChallengesOscillating LossMode CollapseUninformative LossHyperparametersTackling the GAN ChallengesWasserstein GANWasserstein LossThe Lipschitz ConstraintWeight ClippingTraining the WGANAnalysis of the WGANWGAN-GPThe Gradient Penalty LossAnalysis of WGAN-GPSummary
Apples and OrgangesCycleGANYour First CycleGANOverviewThe Generators (U-Net)The DiscriminatorsCompiling the CycleGANTraining the CycleGANAnalysis of the CycleGANCreating a CycleGAN to Paint Like MonetThe Generators (ResNet)Analysis of the CycleGANNeural Style TransferContent LossStyle LossTotal Variance LossRunning the Neural Style TransferAnalysis of the Neural Style Transfer ModelSummary
The Literary Society for Troublesome MiscreantsLong Short-Term Memory NetworksYour First LSTM NetworkTokenizationBuilding the DatasetThe LSTM ArchitectureThe Embedding LayerThe LSTM LayerThe LSTM CellGenerating New TextRNN ExtensionsStacked Recurrent NetworksGated Recurrent UnitsBidirectional CellsEncoder–Decoder ModelsA Question and Answer GeneratorA Question-Answer DatasetModel ArchitectureInferenceModel ResultsSummary
PreliminariesMusical NotationYour First Music-Generating RNNAttentionBuilding an Attention Mechanism in KerasAnalysis of the RNN with AttentionAttention in Encoder–Decoder NetworksGenerating Polyphonic MusicThe Musical OrganYour First MuseGANThe MuseGAN GeneratorChords, Style, Melody, and GrooveThe Bar GeneratorPutting It All TogetherThe CriticAnalysis of the MuseGANSummary

Reinforcement LearningOpenAI GymWorld Model ArchitectureThe Variational AutoencoderThe MDN-RNNThe ControllerSetupTraining Process OverviewCollecting Random Rollout DataTraining the VAEThe VAE ArchitectureExploring the VAECollecting Data to Train the RNNTraining the MDN-RNNThe MDN-RNN ArchitectureSampling the Next z and Reward from the MDN-RNNThe MDN-RNN Loss FunctionTraining the ControllerThe Controller ArchitectureCMA-ESParallelizing CMA-ESOutput from the Controller TrainingIn-Dream TrainingIn-Dream Training the ControllerChallenges of In-Dream TrainingSummary
Five Years of ProgressThe TransformerPositional EncodingMultihead AttentionThe DecoderAnalysis of the TransformerBERTGPT-2MuseNetAdvances in Image GenerationProGANSelf-Attention GAN (SAGAN)BigGANStyleGANApplications of Generative ModelingAI ArtAI Music

Content preview from Generative Deep Learning

Chapter 6. Write

In this chapter we shall explore methods for building generative models on text data. There are several key differences between text and image data that mean that many of the methods that work well for image data are not so readily applicable to text data. In particular:

Text data is composed of discrete chunks (either characters or words), whereas pixels in an image are points in a continuous color spectrum. We can easily make a green pixel more blue, but it is not obvious how we should go about making the word cat more like the word dog, for example. This means we can easily apply backpropagation to image data, as we can calculate the gradient of our loss function with respect to individual pixels to establish the direction in which pixel colors should be changed to minimize the loss. With discrete text data, we can’t apply backpropagation in the usual manner, so we need to find a way around this problem.
Text data has a time dimension but no spatial dimension, whereas image data has two spatial dimensions but no time dimension. The order of words is highly important in text data and words wouldn’t make sense in reverse, whereas images can usually be flipped without affecting the content. Furthermore, there are often long-term sequential dependencies between words that need to be captured by the model: for example, the answer to a question or carrying forward the context of a pronoun. With image data, all pixels can be processed simultaneously.
Text data ...