Fast neural networks implementation in C++

Uncomment required files for the build

./build.sh

Boost C++

sudo apt-get install libboost-all-dev

g++

sudo apt-get install g++

NeuralNetwork nn = NeuralNetwork();
nn.add_layer(new ReluLayer(28 * 28, 128, LEARNING_RATE, LAMBDA));
nn.add_layer(new ReluLayer(128, 32, LEARNING_RATE, LAMBDA));
nn.add_layer(new SoftmaxLayer(32, 10, LEARNING_RATE, LAMBDA));
nn.add_loss_function(new SoftmaxCrossEntropyLoss());

Test the implementation with an interactive jupyter notebook

• class Matrix

  file: deep-learning/matrix/matrix.h

  Supports all basic matrix operations and the ability to do fast matrix dot products.
  All Inputs must be fed into the network using this class.
  All internal calculations work on top of these Matrix objects.
  

  Constructor: Matrix(int n, int m, char rand_init)

  Initialize a n x m matrix with suitable random values.
  rand_init = 'u' for uniform distribution between 0 and 1
  rand_init = 'n' for normal distribution with mean 0 and variance 1
  default rand_init = 0 for zero initialization.
  

• class Layer

  file: deep-learning/layer/layer.h

  The Layer class does what its name says. It represents the layers of the neural
  network. This is an abstract class and implements the key forward propagation,
  backward propagation and optimization steps. Other layers must extend this class
  with its own activation function by overriding the Matrix activation(Matrix Z); and
  Matrix backward_activation(Matrix Z); methods.
  

  Constructor: Layer(int l_, int l, float learning_rate, float lambda, float beta1, float beta2)

  Initialize the layer with the given parameters.
  l_: number of units in previous layer
  l: number of units for the layer
  learning_rate: step size for optimization algorithm
  lambda: regularization hyperparameter
  beta1: first order term for 'adam' optimizer
  beta2: second order term for 'adam' optimizer
  

  The following activations are already implemented:

  1. class SigmoidLayer: public Layer
  2. class ReluLayer: public Layer
  3. class SoftmaxLayer: public Layer

  Mathematical details of the forward and backward propagation steps are described later.

• class LossFunction

  file: deep-learning/loss/loss.h

  This is an abstract base class for all loss functions that can be implemented. Other loss
  functions must implement the float cost(Matrix A, Matrix Y); and
  Matrix derivative(Matrix A, Matrix Y); methods.
  

  The following loss functions are already implemented:

  1. class BinaryCrossEntropyLoss: public LossFunction
  2. class SoftmaxCrossEntropyLoss: public LossFunction

  NOTE: SoftmaxCrossEntropyLoss derivative calculates the derivative of cross entropy loss multiplied with softmax activation layer.

  Mathematical details of the forward and backward propagation steps are described later.

• class NeuralNetwork

  file: deep-learning/neuralnetwork/neuralnetwork.h

  Methods:

  1. void add_layer(Layer * L);
     Adds a fully connected layer to the network.

  2. void add_loss_function(LossFunction * J);
     Adds target loss function to the network.

  3. float train_batch(Matrix X, Matrix Y, int num_iterations);
     Trains weights for num_interations for given input and labels.
     returns cost

  4. Matrix predict(Matrix X);
     

  5. void save(string filename);
     Saves the network and its state to file.

  6. void load(string filename);
     Loads a pre-trained network from file.

NeuralNetwork nn = NeuralNetwork();
nn.add_layer(new SigmoidLayer(NUM_INPUTS, 1, LEARNING_RATE, LAMBDA));
nn.add_loss_function(new BinaryCrossEntropyLoss());
nn.train_batch(X_train, Y_train, 1000);
Matrix Y_pred = nn.predict(X_test);