Neural Network in C++
A neural network implemented with matrices in C++, from scratch !
This program is meant to be used for supervised learning.
What's in there ?
- src/XOR : Learning XOR operation.
- src/XOR_PLOT : Learning XOR operation and plot weights variation on a graph (using python).
- src/Digits-Recognition : Learning to recognize hand-written digits with a training file.
Download, Compile & Run
git clone https://github.com/omaflak/Neural-Network
cd Neural-Network/src
git submodule update --init
# cd into one of the directories above and:
sh compile.sh
./main
Network Class
The Network class contains the gradient descent algorithm.
Both src/XOR and src/Digit-Recognition are using it. Quick description :
// constructor
// vector 'neurons' should contain:
// - number of input neurons at first index
// - number of output neurons at last index
// - number of hidden neurons in between
// example: {2,5,3,1} = 2 input neurons, 1 output neuron, 2 hidden layers (5 neurons and 3 neurons respectively)
// learning rate : experimental
Network(std::vector<int> neurons, double learningRate);
// make prediction
Matrix<double> computeOutput(std::vector<double> input);
// learns from the previous computeOutput()
void learn(std::vector<double> expectedOutput);
// save all network's parameters into a file (after a training)
void saveNetworkParams(const char *filepath);
// load network's parameters from a file so you don't have to train it again
void loadNetworkParams(const char *filepath);
// or use the constructor
Network(const char *filepath);
Plot
I was curious to see what would've happened if I had plotted the network's parameters on a graph, so I did it, and the result is actually fun :)
The program was learning XOR operation and saving it's weights and error variation over time.
Then I plotted the data using plotly
EDIT: I updated the repo and now there is a python script called plotXY.py at src/ that will do the job instead. Try to compile and run src/XOR_PLOT.
And here is the result :
We can see that the program is actually working: while the weights are converging to specific values, the error is decreasing.
Amen !