wrapper

A Java wrapper around Keras

Prerequisites

The following Python packages are required:

• numpy
• Keras

The following Java packages are required:

• com.onpositive.wordnet

Installation

Download a release zip-file and extract it.

Wrapper functionality

• Training your own network;
• Predicting classes for data using trained network

Available elements

• Layers: Dense, BatchNormalization
• Activation functions: relu, softmax
• Algorithms: SGD

Demo example for training

1. Create your own model:

• Create an instance of Model:

Model model = new Model();

• Construct your model by adding layers:

model.add(new Dense(50, "relu", 735));
model.add(new BatchNormalization());
model.add(new Dense(486, "softmax"));

Note that it is needed to indentify an input dimension for the first layer!

• Name your model:

model.setName("MyModel");

• Provide the path, where the model will be saved:

model.setPath("D:/Morph/SimpleTest/model.txt");

2. Create an experiment:

• Create an instance of Experiment with the Model and the way of data preparation:

MyExperiment exp = new MyExperiment(model, new GrammemPreparation());

• Set the parameters for your Experiment:

  exp.setNumberOfClasses(486); // number of target classes
    exp.setTrainingSamples(300); // number of batches (files) with training samples
    exp.setTestSamples(100); // number of batches (files) with validation samples
    exp.setSamplesInFile(5000); // number of samples in batch (file), by default is 1000
  exp.setLearningRate(0.1f) // learning rate, by default is 0.01
  exp.setAlgorithm("Adam") // algorithm for learning, by default is "SGD"

3. Provide the path to the python file "SimpleTest.py":

exp.setPythonFile(YOUR_PATH)

4. Load your data (for this demo you can download this corpus):

List<Sentence> corpus = Corpora.loadCorpus(PATH_TO_DATA, true);

Note that 'true' is only for the corpus on the link above!

5. Now you are ready to run this experiment:

exp.run(corpus, DIRECTORY_WITH_DATA);

Once the experiment finishes, in the directory with your data two new files will be created:

• weights for the trained network;
• statistics for this experiment

Demo example for predicting

1. Create model for weights you have:

• Create an instance of Model:

Model model = new Model();

• Construct your model by adding layers:

model.add(new Dense(50, "relu", 735));
model.add(new BatchNormalization());
model.add(new Dense(486, "softmax"));

Note that it is needed to indentify an input dimension for the first layer!

• Name your model:

model.setName("MyModel");

• Provide the path, where the model will be saved:

model.setPath("D:/Morph/SimpleTest/model.txt");

2. Provide the path to the python file "Prediction.py":

model.setPythonFile(YOUR_PATH);

3. Now you are ready to get predictions:

model.predict(DIRECTORY_WITH_DATA, 3, 5000);

Here in example 3 is the general number of batches (files) you want to get prediction for and 1500 is the general number of samples in a batch (file). For this demo you can use the corpus on the link above.

Make sure that data and weights for the model are located in the same directory!

Once the predictions are made, corresponding files with results will be created in the directory with data.