training_model.md

Training the Character-level CNN Model on a CSV Dataset

This document explains training on an imperative model. To learn about defining an imperative model, see the model definition document.

To train the model as-is, run the following command.

mkdir -p src/main/resources
# Download the dataset
$ wget -O src/main/resources/malicious_url_data.csv https://raw.githubusercontent.com/incertum/cyber-matrix-ai/master/Malicious-URL-Detection-Deep-Learning/data/url_data_mega_deep_learning.csv
# Run train command
$ ./gradlew train

The default training runs for 7 epochs. Pre-set training configurations are discussed in-detail in the following section.

Training configuration

Deep Java Library (DJL) provides a TrainingConfig class to define hyperparameters for training. Hyperparameters, including learning rate and optimizer, are provided to the Trainer object using the TrainingConfig class.

For example, this trainingConfiguration is used to initialize the Trainer object, which in turn initializes the parameters of the model.

    /**
    *Learning Rate definition, FactorTracker is a parent for similar hyperparameters
    * Sets learningRate and number of steps
    */
    int learningRate = 0.01;
    int stepSize = inputDataSize/batchSize;
    FactorTracker factorTracker =
            LearningRateTracker.factorTracker()
                    .optBaseLearningRate(learningRate)
                    .setStep(stepSize)
                    .build();
    // Setting optimizer object, in this case build an SGD optimizer with momentum
    Optimizer optimizer =
            Optimizer.sgd()
                    .setRescaleGrad(1.0f / batchSize)
                    .setLearningRateTracker(factorTracker)
                    .optWeightDecays(0.00001f)
                    .optMomentum(0.9f)
                    .build();

    // Define loss and Initializer 
     Loss loss = Loss.softmaxCrossEntropyLoss();
     Initializer initializer = new XavierInitializer(
                                XavierInitializer.RandomType.UNIFORM,
                                XavierInitializer.FactorType.AVG,
                                2.24);
     //Set distribution of Initializer Randamozier to uniform and Factor to be average, with magnitude.
     
    //Use the above to create a TrainingConfig object
  
    TrainingConfig trainingConfig = new DefaultTrainerConfig(initializer, loss)
                                       .setOptimizer(optimizer)
                                        .setBatchSize(batchSize)
                                       .setDevices(new Device[] {Device.defaultDevice()});

Trainer

The Trainer is a per-model instance object that provides training functionality using a simple API.

The following commands initialize and use trainer:

//Define and load model
Model model = Model.newInstance();
model.load(modelName, modelName);
//create a Trainer with the Training Config
Trainer trainer = model.newTrainer(trainingConfig);


//initialize the parameters , pass shape of input.
trainer.initialize(inputShape);

//train on dataSet for epochs

for (int epoch = 0; epoch < 10; epoch++) {
   for (Batch batch: trainer.iterateDataset(trainDataset) { // trainDataset is a Dataset Object, containing TRAIN split
       trainer.trainBatch(batch);
       trainer.step();
       batch.close();
    }
   //Validate batch with Validate split
    for (Batch batch : trainer.iterateDataset(validateDataset)) {
        trainer.validateBatch(batch);
        batch.close();
    }
    // Save Model after current epoch
    Model model = trainer.getModel();
    model.setProperty("Epoch", String.valueOf(epoch));
    model.save(Paths.get(outputDir), "modelName");
}

The Trainer object manages the lifecycle of training from initialization to backPropagation on the model. The Trainer handles loading the dataset and batching during training.