This is Deep neural Network written in Java from scratch based on Andrew Ng Coursera course(https://www.youtube.com/channel/UCcIXc5mJsHVYTZR1maL5l9w). Only basic l-layered neural network is developed. This project implements starting from 2- layer network to L-layer generic framework. By generic I mean the same functions can be called for regression and classification problem by just putting the right parameters such as final layer activation function, loss etc. This project is to help students for doing research in the deep learning framework such as parameter initialization, new activation function, implementing drop-outs. I have also added the code for the GPU using jCuda library. Further I have given file on disc creation operation, if some one lacks RAM, and GPU. Although this is not completely implemented but initial large matrix object creation and accessing, multiplication is implemented directly from the disc. One can easily implement the deep neural network process, thanks to the JAVA fast input output libraries performance is also considerable for large scale machine learning task. There is TestMain.Java class for three kinds of examples i.e regression, classification and multiclass classificaiton. Multi label classification can also be achieved as onVsAll method is also implemented. This project focusses on working on taskes which are not easily available such as devolping your own method for parameter initilization, writing your own activation, changing neural network architecture etc. If you use this code please cite the following paper:
- Dong, Shi, Ping Wang, and Khushnood Abbas. " A survey on deep learning and its applications." Computer Science Review 40 (2021): 100379.
#Examples
import java.util.Arrays;
import java.util.Map;
import com.jnn.consts.IPredictionModelConstants;
import com.jnn.enums.DeepNNActivations;
import com.jnn.enums.DeepNNCostFunctions;
import com.jnn.enums.DeepNNKeysForMaps;
import com.jnn.enums.DeepNNOptimizers;
import com.jnn.enums.DeepNNRegularizor;
import com.jnn.enums.ParameterInitializationType;
import com.jnn.utilities.DNNUtils;
import com.jnn.utilities.MatrixUtil;
import com.jnn.utilities.Stopwatch;
import com.phd.models.NeuralNetworkModels;
String[] files= {"BostonHousePrinceDataset\\\\regressionBostonHousePrince.csv",
"ionosphereDataset\\ClassificatinoDataSet.txt",
"heartDisease\\heart.txt",
"CourseraHandWrittenDigits\\handWrittenDigitsMultiClassClassification.txt",
};
String path=".\\data\\"+files[2];//0 for regression, 1,2 Binary classification, 3 for multiclass classificaiton
//It is supposed that last column in csv file will be the label column
double[][] X_train = null;
double[][] Y_train = null;
double[][] X_test = null;
double[][] Y_test = null;
int m;
int numofOfHiddenLayers=3;
int nodesInHiddenLayer = 50;
double learningRate=0.0075;
double trainRatio=0.7;
boolean isClassification=true;/// This is needed for random sampling, because for regression problem examples can be sampled uniformly
//for training and testing while for classificaiton we need to sample examples from all the classes.
Map<Enum<DeepNNKeysForMaps>, Object> resultMap=DNNUtils.readFileAsMatrix(path, ",",Boolean.FALSE,trainRatio,isClassification);
X_train=(double[][])resultMap.get(DeepNNKeysForMaps.KEY_FOR_FEATURE_MATRIX);
Y_train=(double[][])resultMap.get(DeepNNKeysForMaps.KEY_FOR_OUTPUT_LABEL);
X_test=(double[][])resultMap.get(DeepNNKeysForMaps.KEY_FOR_FEATURE_MATRIX_TEST);
Y_test=(double[][])resultMap.get(DeepNNKeysForMaps.KEY_FOR_OUTPUT_LABEL_TEST);
m=(int)resultMap.get(DeepNNKeysForMaps.KEY_FOR_NUMBER_EXAMPLES);
int noOfBatches=X_train.length/500;
if(noOfBatches<1) {
noOfBatches=1;
}
int numberOfDimenstions=(int)resultMap.get(DeepNNKeysForMaps.KEY_FOR_NUMBEROFDIMENSIONS);
Map<String, Object> learnedParameters=null;
Map<DeepNNKeysForMaps, Object> oneHotMatrixresult=null;
X_test=MatrixUtil.T(X_test);
Y_test=MatrixUtil.T(Y_test);
int numberOfClasses=1;
if(isClassification) {
numberOfClasses=(int)resultMap.get(DeepNNKeysForMaps.KEY_FOR_NUMBER_OF_CLASSES);;
}
int type=1;//1: Multiclass classification, 2: binary classification , 3: Regression problem
MatrixUtil.setSeed(IPredictionModelConstants.RANDOM_SEED_FOR_REPRODUCIBILITY);
MatrixUtil.setGpuEnabled(false);
int[] layerDimenstions= new int[numofOfHiddenLayers+2];
for (int i=1;i<=(numofOfHiddenLayers);i++) {
layerDimenstions[i]=(int)(nodesInHiddenLayer/( 1));
if(layerDimenstions[i]<1) {
layerDimenstions[i]=numberOfDimenstions;
}
}
layerDimenstions[0]=numberOfDimenstions;
layerDimenstions[numofOfHiddenLayers+1]=numberOfClasses;
Map<Double,Map<String, Object>> multiClassClassificationParameters=null;
System.out.println("Training started");
if(type==1) {
Stopwatch stopWatch = new Stopwatch();
X_train=MatrixUtil.divide(X_train, MatrixUtil.max(X_train));
X_test=MatrixUtil.divide(X_test, MatrixUtil.max(X_test));
System.out.println(" time taken: " + stopWatch.elapsedTime()
+ " seconds \n");
DeepNNActivations hiddenLayerActivations=DeepNNActivations.TANH;
ParameterInitializationType paramInitType=ParameterInitializationType.XAVIOUR;
learnedParameters = NeuralNetworkModels.l_LAYER_MODEL_Generic(X_train, Y_train, layerDimenstions,paramInitType, learningRate, 1000, DeepNNOptimizers.ADAM,DeepNNRegularizor.DROP_OUT,0.80,hiddenLayerActivations, DeepNNActivations.SOFTMAX,DeepNNCostFunctions.CATEGORICAL_CROSS_ENTROPY,noOfBatches,true, true);
String dropOut=paramInitType+"_dropout_82_";
oneHotMatrixresult=MatrixUtil.createOneHotVectorMatrix(MatrixUtil.T(Y_test));
double[][] Y_OneHot=(double[][])oneHotMatrixresult.get(DeepNNKeysForMaps.KEY_FOR_ONE_HOT_MATRIX);
Map<Double, double[]> labeltoOnhotVectorMap=(Map<Double, double[]>)oneHotMatrixresult.get(DeepNNKeysForMaps.KEY_FOR_LABEL_TO_ONE_HOT);
double[][] predictedAL=DNNUtils.getPredictedResult_multiclass_classification(X_test, learnedParameters);
Y_OneHot=MatrixUtil.T(Y_OneHot);
double finalPrecision = DNNUtils.accuracyForMultiClassClassification(predictedAL, Y_OneHot);
MatrixUtil.print("Predicted multiclass classification on test data precision= " + finalPrecision);
}
else if (type==2) {
//learnedParameters=NeuralNetworkModels.l_LAYER_MODEL_BINARY_CLASSIFICATION_WithDropOut(X_train, Y_train, layerDimenstions, learningRate, 5000,DeepNNOptimizers.GD, Boolean.TRUE,.95);
learnedParameters=NeuralNetworkModels.l_LAYER_MODEL_BINARY_CLASSIFICATION(X_train, Y_train, layerDimenstions, learningRate, 5000,DeepNNOptimizers.ADAM, Boolean.TRUE);
double[][] predictedAL=DNNUtils.getPredictedActivationScoreForGenericActivations(X_test, learnedParameters,DeepNNActivations.TANH, DeepNNActivations.SIGMOID);
MatrixUtil.print("Predicted precision on test dataset= " + DNNUtils.accuracyForClassification(predictedAL, Y_test));
}
else if (type==3) {
learnedParameters=NeuralNetworkModels.l_LAYER_MODEL_Regression(X_train, Y_train, layerDimenstions, learningRate, 10000,DeepNNOptimizers.ADAM, Boolean.TRUE);
double[][] predictedAL=DNNUtils.getPredictedResult_regression(X_test, learnedParameters);
MatrixUtil.print("Predictions on Test data= " + Arrays.deepToString(predictedAL));
MatrixUtil.print("Real on Test data= " + Arrays.deepToString(Y_test));
}