README.md

Eclipse Deeplearning4J Examples

This project contains a set of examples that demonstrate use of the high level DL4J API to build a variety of neural networks. The DL4J ecosystem also allows users to build neural networks with SameDiff (part of the ND4J library) with a more fine grained API. More information on that can be found here

The pom file in this project can be used as a template for a user's own project. The examples in this project and what they demonstrate are briefly described below. This is also the recommended order to explore them in.

Go back to the main repository page to explore other features/functionality of the Eclipse Deeplearning4J ecosystem. File an issue here to request new features.

QUICKSTART

Modeling Examples

Feedforward Neural Networks

Classification

• IrisClassifier.java Basic end to end example that introduces important concepts like RecordReaders, MultiLayerConfiguration etc
• LinearDataClassifier.java Basic end to end example with plots
• MNISTSingleLayer.java Classify MNIST
• MNISTDoubleLayer.java Classify MNIST with more layers
• ModelXOR.java Model a two input XOR function (ie. a simple non linearly separable function)
• MoonClassifier.java Model data that separates into a "moon" shape and visualize the results
• SaturnClassifier.java Model data that separates into a "saturn" shape and visualize the results

Regression

• CSVDataModel.java Basic end to end example with plots
• MathFunctionsModel.java Model various mathematical functions
• SumModel.java Model addition on a synthetic dataset with noise
• ImageDrawer.java Train a model to draw an image!

Unsupervised

• MNISTAutoencoder.java A basic introduction to how to build an autoencoder

Convolutional Neural Networks

• LeNetMNIST.java The classic LeNet example for classifying hand-drawn digits (MNIST)
• LeNetMNISTReLu.java The same as above with minor modifications
• CIFARClassifier.java Classify the CIFAR dataset
• CenterLossLeNetMNIST.java Train an embedding using the center loss model, on MNIST

Recurrent Neural Networks

• UCISequenceClassification.java Time series (sequence) classification on the UCI syntetic control time series dataset
• MemorizeSequence.java Train a RNN to memorize a sequence of characters
• RNNEmbedding.java Use an EmbeddingLayer (equivalent to using a DenseLayer with a one-hot representation for the input) as the first layer in an RNN
• VideoFrameClassifier.java Classify shapes that appear in a video frame. Demonstrates how to combine RNN, CNN and fully connected dense layers in a single neural network. This is a memory consuming example. You need at least 7G of off heap memory. Refer here to configure memory off heap.

Variational Auto Encoder

• VaeMNISTAnomaly.java Unsupervised anomaly detection on MNIST using a variational autoencoder
• VaeMNIST2dPlots.java Train a variational autoencoder on MNIST and plot MNIST digit reconstructions vs. the latent space as well as the latent space values for the MNIST test set as training progresses

Features

• SaveLoadMultiLayerNetwork.java Save and load a multilayer neural network
• SaveLoadComputationGraph.java Save and load a computational graph
• EarlyStoppingMNIST.java Early stopping example on a subset of MNIST i.e conduct training and use the parameters that give the minimum test set loss
• PreSaveFirst.java & LoadPreSavedLenetMnistExample.java Save time when training with multiple epochs by presaving datasets
• WeightedLossFunctionExample.java Out of the box loss function that can be used with imbalanced classes
• BasicUIExample.java Basic UI example
• UIStorageExample.java Save training data to a file and reload it later to display in the UI
• RemoteUIExample.java If you need the UI to be hosted in a separate JVM for training.
• TSNEStandardExample.java Basic TSNE

ADVANCED

Modeling Examples

Computer Vision

• TinyYoloHouseNumberDetection.java Transfer learning from a Tiny YOLO model pretrained on ImageNet and Pascal VOC to perform object detection with bounding boxes on The Street View House Numbers Dataset.
• NeuralStyleTransfer.java Neural Style Transfer Algorithm
• MultiDigitNumberRecognition.java Captcha recognition

Natural Language Processing

Text Classification

With pretrained word2vec:

• ImdbReviewClassificationRNN.java Sentiment Classification on the IMDB dataset with a RNN model
• ImdbReviewClassificationCNN.java Sentiment Classification on the IMDB dataset with a CNN model

Generating Embeddings:

• Paragraph Vectors
• Sequence Vectors
• Word2Vec

Modeling with a word2vec model trained on a custom corpus:

• PrepareWordVector.java, TrainNews.java Sentence classification using a word2vec model training on a custom corpus

Char Modelling

• GenerateTxtModel.java & GenerateTxtCharCompGraphModel.java MultiLayerNetwork and ComputationGraph versions of a model that is trained to "write Shakespeare" one character at a time, inspired by Andrej Karpathy's now famous blog post.

Other Sequence Modeling Examples

• SequenceAnomalyDetection.java Anomaly detection on sequence sensor data
• TrainLotteryModelSeqPrediction.java Model trained on a synthetic dataset that attempts to uncover the contrived pattern.

Specific Models and Special Architectures

• AlphaGoZeroTrainer.java Train AlphaGo Zero model on dummy data.
• DenseNetMain.java Builds dense net to classify a small set of animal images. Augments the dataset with transforms like blur etc.
• AdditionModelWithSeq2Seq.java A seq2seq model that learns to add

Features

Customizing DL4J

• CustomActivationUsageEx.java Implement custom activation functions
• CustomLayerUsageEx.java Implement custom layers
• CustomLossUsageEx.java Implement custom loss functions

NOTE: SameDiff which is part of ND4J gives users a way to customize DL4J. More information on that is found here

Performance

• ParallelInferenceExample.java How to run parallel inference in DL4J

Debugging

• CSVExampleEvaluationMetaData.java Trace where data from each example comes from and get metadata on prediction errors
• MultiLayerNetworkExternalErrors.java Train a MultiLayerNetwork where the errors come from an external source, instead of using an Output layer and a labels array.

TransferLearning

Demonstrates use of the dl4j transfer learning API which allows users to construct a model based off an existing model by modifying the architecture, freezing certain parts selectively and then fine tuning parameters. Read the documentation for the Transfer Learning API at https://deeplearning4j.konduit.ai/tuning-and-training/transfer-learning.

• EditLastLayerOthersFrozen.java Modifies just the last layer in vgg16, freezes the rest and trains the network on the flower dataset.
• FeaturizedPreSave.java & FitFromFeaturized.java Save time on the forward pass during multiple epochs by "featurizing" the datasets. FeaturizedPreSave saves the output at the last frozen layer and FitFromFeaturize fits to the presaved data so you can iterate quicker with different learning parameters.
• EditAtBottleneckOthersFrozen.java A more complex example of modifying model architecure by adding/removing vertices
• FineTuneFromBlockFour.java Reads in a saved model (training information and all) and fine tunes it by overriding its training information with what is specified