A small and easy to use neural net implementation for C++. Just download and #include!

Last update: Nov 11, 2022

Related tags

Overview

NN++

A short, self-contained, and easy-to-use neural net implementation for C++. It includes the neural net implementation and a Matrix class for basic linear algebra operations. This project is mostly for learning purposes, but preliminary testing results over the MNIST dataset show some promise.

Getting Started

These instructions will get you a copy of the net up and running on your local machine for development and testing purposes.

Prerequisites

Any compiler that can handle C++11.

Installing

Download Matrix.hpp, Matrix.cpp, NeuralNet.hpp, and NeuralNet.cpp and place them in your project's working directory.
Include the headers in your main driver program (e.g. main.cpp).

#include "Matrix.hpp"
#include "NeuralNet.hpp"

NOTE: It is not required to #include "Matrix.hpp" since it is included within NeuralNet.hpp. However it is probably better to do so for clarity and safety in case you plan on using Matrix objects in your code (and you probably will if you use NeuralNet).

Example Code

The Matrix Class

First you need to know how to use the Matrix class. Matrix objects are basically 2D-vectors with built-in linear algebra operations.

Matrix Initialization

Matrix A;       // Initializes a 0x0 matrix.
Matrix B(2,2)   // Initializes a 2x2 matrix with all zeros. Values are doubles.
Matrix C(2,1)   // Initializes a 2x1 matrix.

Element Access

To access/modify a value in a matrix, use operator(), NOT operator[]:

B(0,0) = 1; B(0,1) = 2; B(1,0) = 3; B(1,1) = 4;   // [1    2]
                                                  // [3    4]

C(0,0) = 1; C(1,0) = 2;                           // [1]
                                                  // [2]

Matrix Term-by-Term Addition/Subtraction/Multiplication

// Commutative property is supported for addition
Matrix D = B+B;       // D = [2   4]
                             [6   8]
                             
Matrix E = B-B;       // E = [0   0]
                             [0   0]
                             
// Commutative property is supported for multiplication                             
Matrix F = B*B        // F = [1   4]
                             [9  16]
                             
// Mismatching matrix dimensions in term-by-term operations
// is illegal and a MatrixDimensionsMismatch exception will be thrown.
Matrix G = B+C;       // Throws MatrixDimensionsMismatch()
Matrix G = B-C;       // Throws MatrixDimensionsMismatch()
Matrix G = B*C;       // Throws MatrixDimensionsMismatch()

Matrix and Scalars

// Commutative property is supported for addition
Matrix BplusTwo = B+2;  // (== 2+B)   BplusTwo = [3   4]
                                                 [5   6]

Matrix CminusTwo = C-2; //           CminusTwo = [-1]
                                                 [ 0]
                                                 
Matrix TwominusB = 2-C; //           TwominusB = [ 1]
                                                 [ 0]
                                                 
// Commutative property is supported for multiplication
Matrix BtimesThree = B*3; // (== 3*B) BtimesThree = [3    6]
                                                    [9   12]

Matrix Multiplication (Dot Product)

Matrix BB = B.dot(B);     // BB = [ 7  10]
                                  [15  22]
                                  
Matrix BC = B.dot(C);     // BC = [ 5]
                                  [11]

// Mismatching the number of columns in the left-hand-side matrix
// with the number of rows in the right-hand-side matrix is illegal
// A MatrixInnderDimensionsMismatch exception will be thrown.
Matrix CB = C.dot(B);     // Throws MatrixInnderDimensionsMismatch()

Matrix Transpose

Matrix B_T = B.T();   // B_T = [1   3]
                               [2   4]
                                 
Matrix C_T = C.T();   // C_T = [1   2]

An Example of Populating a 4x3 Matrix

int m = 4;
int n = 3;

Matrix mtrx(m,n);

int count = 1;
for (int i = 0; i < mtrx.getNumOfRows(); ++i) {
    for (int j = 0; j < mtrx.getNumOfCols(); ++j) {
        mtrx(i,j) = count;
        ++count;
    }
}

This will result with mtrx ==

[ 1     2     3]
[ 4     5     6]
[ 7     8     9]
[10    11    12]

The NeuralNet Class

Neural Net Initialization (The Parameters)

When initialized, a net takes in five parameters:

Number of input nodes.
Number of nodes per hidden layer.
Number of output nodes.
Number of hidden layers.
The learning rate.

NeuralNet NN(4, 3, 1, 10, 0.1);

This particular neural net has 4 input nodes, 1 hidden layer with 3 nodes, 10 output node, and has a learning rate of 0.1.
New neural nets' weights are initialized with values drawn from a normal distribution centered at 0, with standard deviation that is equal to 1/sqrt(number_of_inputs_to_nodes_in_next_layer). In other words, small negative and positive values that are proportional to the size of their previous layer.

A Training Cycle

Once the net is initialized, it is ready to do work.
ONE training cycle == one feed forward and one back propagation with weight adjustments.

To train one cycle, the input data must be parsed into a Matrix object with dimensions: 1xnumber_of_input_nodes (1x4 in our case), and the target output must be parsed into a Matrix object with dimensions: 1xnumber_of_output_nodes (1x10 in our case).

Matrix input(1,4);
input(0,0) =  0.3;
input(0,1) = -0.1;
input(0,2) =  0.2;
input(0,3) =  0.8;

Matrix targetOutput(1,1);
target(0,0) =  0.5;
target(0,1) = -0.3;
        .
        .
        .
target(0,9) =  0.23;        // Obviously, matrices should be populated using
                            // some parser and not manualy like this.

Then, simply execute the training cycle on the data as follows:

NN.trainingCycle(input, targetOutput);

Repeate the process over all training instances.

Querying the Net

Once the training phase is complete, you can query it as follows:
(Technically speaking, you can query it right after initialization).

Parse the query into a Matrix like parsed the training instance:

Matrix query(1,2);
input(0,0) =  0.5;
input(0,1) = -0.2;
input(0,2) = -0.3;
input(0,3) =  0.4;

Query the net and catch the result:

Matrix prediction = NN.queryNet(query);   // Will return a 1x10 Matrix object with net's prediction

AND THAT'S IT!

TODO

Add array, std::vector, and std::initializer_list constructors to the Matrix class
Either improve on or replace my Matrix class for better/faster performance
Add multiple epoch learning with early stopping.

Authors

Gil Dekel - Initial implementation - stagadish

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

You might also like...

Inject .NET assemblies into an existing process

inject-assembly - Execute .NET in an Existing Process This tool is an alternative to traditional fork and run execution for Cobalt Strike. The loader

385 Dec 30, 2022

Mirror of the HPSx64 Repository on Sourceforge.net

Full Program Name: Highly-Experimental Playstation Simulator x64 Program Author: TheGangster Platforms: Windows 64-bit Contents: hps1x64 - Playstat

5 Sep 12, 2022

C++11 header-only library that offers small vector, small flat map/set/multimap/multiset.

sfl library This is header-only C++11 library that offers several new containers: small_vector small_flat_set small_flat_map small_flat_multiset small

21 Dec 14, 2022

Fast and easy to use, high frequency trading framework for betfair

Hedg Fast and easy to use, high frequency trading framework for betfair About Hedg In the sports trading industry, low latency is really important. Th

8 Jun 11, 2022

Generalized easy to use SPI library for Arduino designed mainly for immense data transfer through multiple slaves. This library features Cyclic Redundancy Check (CRC) and user defined data transfer which inspired by Nick Gammon.

SPI Master and Slave Library If you don't have VS Code you can install from here. After downloading click extensions and type and install PlatformIO I

1 Oct 6, 2021

Gfx - A minimalist and easy to use graphics API.

gfx gfx is a minimalist and easy to use graphics API built on top of Direct3D12/HLSL intended for rapid prototyping. It supports: Full shader reflecti

343 Dec 29, 2022

MasterPlan is a project management software / visual idea board software. It attempts to be easy to use, lightweight, and fun.

MasterPlan is a customizeable graphical project management software for independent users or small teams. If you need to share plans across a whole co

444 Dec 23, 2022

A simple and easy-to-use Lua library to enjoy videogames programming

raylib-lua-sol Lua bindings for raylib, a simple and easy-to-use library to enjoy videogames programming, with sol (www.raylib.com) raylib-lua-sol bin

65 Dec 6, 2022

The repository contains some examples of pre-trained SNN (Spiking Neural Network) models.

About the Project The repository contains some examples of pre-trained SNN (Spiking Neural Network) models. The models were trained using the MM-BP tr

13 Nov 24, 2022

Comments

Add some work on operator << and a (very) general training function

Happy to do some more work to get this up to scratch (I'm new to performance heavy c++).

I started work on a training function that takes a std::vector<std::pair<Matrix, Matrix>> which and a lambda (this is for defining your own end conditions) should make training simple.

I can add some documentation on how to use it you like it.

Thanks

opened by Cypher1 6

Some issues regarding your library

Hey,

Today I have got to know about your library thru your post in G+. I am also trying to write a similar library/toolbox to be used in optimization.

At first look at your library, I can say that it looks neat. However, it would be great if you considered using templates in your data structures. Or, if you would like to go for inheritance for some reason, then you can abstract some functionality in your base classes.

What I mean by the above comment is that, for example, you could use an activation function abstraction and keep a pointer inside your NeuralNet class to support different activation functions, such as:

#include <iostream>
#include <vector>

template <typename T>
class IFunction {
public:
  virtual T operator()(T variable) const = 0;
  virtual ~IFunction() = default;
};

template <typename T>
class PReLU : public IFunction<T> {
private:
  T alpha_;
public:
  PReLU(T alpha = T { 0 }) : alpha_ {alpha} { }
  T operator()(T variable) const override {
    T zero { 0 };
    return variable < zero ? alpha_*variable : variable;
  }
};

template <typename T>
class Identity : public IFunction<T> {
public:
  T operator()(T variable) const override {
    return variable;
  }
};

int main (int argc, char* argv[]) {
  Identity    <double>  identity_function {         };
  PReLU       <double>  relu_function     {         };
  PReLU       <double>  absolute_value    { -1.     };

  IFunction   <double>  *generic_func_ptr { nullptr };

  size_t numelems { 100ul };
  double xmin { -5. }, xmax { 5. }, dx {(xmax - xmin)/(numelems-1)};
  std::vector <double>  xvalues ( numelems );
  for (size_t idx = 0; idx < xvalues.size(); idx++)
    xvalues[idx] = xmin + idx*dx;

  // either directly call as if your instances were functions (actually, they
  // are functors right now, with their operator()'s overloaded)
  for (auto x : xvalues)
    std::cout << "x: " << x
              << ", identity_function(x): " << identity_function(x)
              << ", relu_function(x): " << relu_function(x)
              << std::endl;

  // or, use polymorphism
  generic_func_ptr = &identity_function;
  for (auto x : xvalues)
    std::cout << "x: " << x
              << ", generic_func_ptr->operator()(x): "
              << generic_func_ptr->operator()(x)
              << std::endl;

  // or, both
  generic_func_ptr = &relu_function;
  for (auto x : xvalues)
    std::cout << "x: " << x
              << ", (*generic_func_ptr)(x): " << (*generic_func_ptr)(x)
              << std::endl;

  // absolute value function
  for (auto x : xvalues)
    std::cout << "x: " << x
              << ", absolute_value(x): " << absolute_value(x)
              << std::endl;

  return 0;
}

I hope this helps :) Good luck with your code!...

opened by aytekinar 3