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.
These instructions will get you a copy of the net up and running on your local machine for development and testing purposes.
Any compiler that can handle C++11.
- Download
Matrix.hpp,Matrix.cpp,NeuralNet.hpp, andNeuralNet.cppand 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).
First you need to know how to use the Matrix class. Matrix objects are basically 2D-vectors with built-in linear algebra operations.
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.
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]
// 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()
// 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 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 B_T = B.T(); // B_T = [1 3]
[2 4]
Matrix C_T = C.T(); // C_T = [1 2]
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]
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.
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.
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!
- Add
array,std::vector, andstd::initializer_listconstructors to the Matrix class - Either improve on or replace my Matrix class for better/faster performance
- Add multiple epoch learning with early stopping.
- Gil Dekel - Initial implementation - stagadish
See also the list of contributors who participated in this project.
This project is licensed under the MIT License - see the LICENSE.md file for details