private static int getDatasetFeaturesCount( String dataset_path ) {

try {

String delimiter = "\t";

File file = new File(dataset_path);

InputStream fis = new FileInputStream(file);

InputStreamReader isr = new InputStreamReader(fis, Charset.forName("UTF-8"));

BufferedReader br = null;

try

{

br = new BufferedReader(isr);

String line0 = br.readLine();

int input_size = line0.split(delimiter).length - 1;

return input_size;

}

finally

{

br.close();

}

}

catch(Exception ex) {

System.out.println("Could not load file "+dataset_path );

System.exit(1);

return -1;

}

}

private static DataSetIterator loadDataset( String dataset_path, int input_size, int batch_size )

{

try {

System.out.format("Loading dataset from %s...", dataset_path);

char delimiter = '\t';

File file = new File(dataset_path);

int real_batch_size = batch_size;

if( real_batch_size<=0 ) {

LineNumberReader lineNumberReader = new LineNumberReader(new FileReader(file));

lineNumberReader.skip(Long.MAX_VALUE);

real_batch_size = lineNumberReader.getLineNumber();

}

int numLinesToSkip = 0;

RecordReader recordReader = new CSVRecordReader(numLinesToSkip,delimiter);

recordReader.initialize(new FileSplit(file));

int labelIndex = input_size;

int numClasses = 2;

DataSetIterator iterator = new RecordReaderDataSetIterator(recordReader,real_batch_size,labelIndex,numClasses);

System.out.println("done");

return iterator;

}

catch(Exception ex)

{

System.out.println("Could not load file "+dataset_path );

System.exit(1);

return null;

}

}

private static MultiLayerNetwork createModel(int input_size)

{

int numOutputs = 2;

int numHiddenNodes = input_size;

int seed=123456;

double learning_rate = 0.05;

MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()

.seed(seed)

.iterations(1)

.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)

.learningRate(learning_rate)

.updater(Updater.NESTEROVS) //To configure: .updater(new Nesterovs(0.9))

.list()

.layer(0, new DenseLayer.Builder().nIn(input_size).nOut(numHiddenNodes)

.weightInit(WeightInit.XAVIER)

.activation(Activation.RELU)

.build())

.layer(1, new DenseLayer.Builder().nIn(numHiddenNodes).nOut(numHiddenNodes/2)

.weightInit(WeightInit.XAVIER)

.activation(Activation.RELU)

.build())

.layer(2, new DenseLayer.Builder().nIn(numHiddenNodes/2).nOut(numHiddenNodes/3)

.weightInit(WeightInit.XAVIER)

.activation(Activation.RELU)

.build())

.layer(3, new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)

.weightInit(WeightInit.XAVIER)

.activation(Activation.SOFTMAX).weightInit(WeightInit.XAVIER)

.nIn(numHiddenNodes/3).nOut(numOutputs).build())

.pretrain(false).backprop(true).build();

MultiLayerNetwork model = new MultiLayerNetwork(conf);

model.init();

model.setListeners(new ScoreIterationListener(10)); //Print score every 10 parameter updates

return model;

}

public static void main(String[] args) {

int batch_size = 100;

String train_path = "/home/eek/polygon/WordRepresentations/data/Xy_train.csv";

int input_size = getDatasetFeaturesCount(train_path);

DataSetIterator train_iter = loadDataset(train_path, input_size, batch_size);

DataSetIterator val_iter = loadDataset("/home/eek/polygon/WordRepresentations/data/Xy_val.csv", input_size,0);

DataSet val_ds = val_iter.next();

INDArray x_val = val_ds.getFeatureMatrix();

INDArray y_val = val_ds.getLabels();

int numOutputs = 2;

int numHiddenNodes = input_size;

MultiLayerNetwork model = createModel(input_size);

String checkpoint_file = "/home/eek/polygon/WordRepresentations/data/deeplearning4j.model";

int nEpochs = 50;

double best_val_acc = 0.0;

int nb_no_improvements = 0;

for ( int iter = 0; iter < nEpochs; iter++) {

System.out.format("Start iteration #%d\n", iter);

model.fit( train_iter );

System.out.println("Evaluating...");

INDArray output = model.output(x_val);

//System.out.println(output);

// let Evaluation prints stats how often the right output had the

// highest value

Evaluation eval = new Evaluation(2);

eval.eval(y_val, output);

System.out.println(eval.stats());

double val_acc = eval.accuracy();

if( val_acc>best_val_acc ) {

best_val_acc = val_acc;

nb_no_improvements = 0;

System.out.format("New best val_acc=%s\n", best_val_acc);

try {

File locationToSave = new File(checkpoint_file);

boolean saveUpdater = false;

ModelSerializer.writeModel(model, locationToSave, saveUpdater);

System.out.format("Model saved in %s\n", checkpoint_file);

}

catch(java.io.IOException ex) {

System.out.println("Error occured when saving model to file: "+ex.getMessage());

System.exit(-1);

}

}

else

{

nb_no_improvements++;

if( nb_no_improvements>10) {

System.out.format("Early stopping occured after not improving val_acc during %d epochs\n", nb_no_improvements);

break;

}

}

}

MultiLayerNetwork restored_model = null;

try {

// Загрузим оптимальный вариант весов сетки.

restored_model = ModelSerializer.restoreMultiLayerNetwork(checkpoint_file);

}

catch(java.io.IOException ex) {

System.out.println("Error occured when saving model to file: "+ex.getMessage());

System.exit(-1);

}

System.out.println("Final evaluation with holdout dataset");

DataSetIterator holdout_iter = loadDataset("/home/eek/polygon/WordRepresentations/data/Xy_holdout.csv", input_size, 0);

DataSet holdout_ds = holdout_iter.next();

INDArray x_holdout = holdout_ds.getFeatureMatrix();

INDArray y_holdout = holdout_ds.getLabels();

INDArray output = restored_model.output(x_holdout);

Evaluation eval = new Evaluation(2);

eval.eval(y_holdout, output);

System.out.println(eval.stats());

return;

}