NeuralGS is a neural network built in c++ with eigen3 as the underlying linear algebra libary. The network has no gpu accelerated methods as of yet. The network makes use of OpenMP to multithread heavy functions.
Creating simple neural network model.
Loss *loss = new CategoricalCrossentropy();
Optimizer *adam = new Adam(0.001, 1e-5, 1e-7, 0.9, 0.999);
Model model;
model.setLoss(loss);
model.setOptimizer(adam);
Layer *firstLayer = new Layer(30, 128, 0, 5e-4, 0, 5e-4);
Activator *relu = new ReLu();
Layer *secondLayer = new Layer(128, 128, 0, 5e-4, 0, 5e-4);
Activator *reluTwo = new ReLu();
Layer *thirdLayer = new Layer(128, 128, 0, 5e-4, 0, 5e-4);
Activator *reluThree = new ReLu();
Layer *fourtLayer = new Layer(128, 325, 0, 0, 0, 0);
Activator *smax = new Softmax();
model.appendLayer(firstLayer, relu);
model.appendLayer(secondLayer, reluTwo);
model.appendLayer(thirdLayer, reluThree);
model.appendLayer(fourtLayer, smax);Training models, where dataset and labels are loaded in to the variabels x and y.
Eigen::MatrixXd x;
Eigen::VectorXi y;
int epochs = 10;
int printGap = 1;
model.train(x, y, epochs, printGap);Saving and loading models
model.train(x, y, epochs, printGap);
model.saveModel();
//Loading
json savedFile
Model model;
model.loadModel(savedFile);Testing and predicting with model
Eigen::MatrixXd testX;
Eigen::VectorXi testY;
model.test(x, y);
//Predicting
sampleInput = Eigen::MatrixXd(3,3);
sampleInput << 0.2,0.3,0.4,
0.5,0.7,0.8,
0.4,0.5,0.6;
model.evaluate(sampleInput)Eigen3 - Linear algebra Nhloman::json - Saving and loading models OpenMP - Multithreading