Skip to content

Latest commit

 

History

History
219 lines (144 loc) · 7.08 KB

README.md

File metadata and controls

219 lines (144 loc) · 7.08 KB

Real Neural Network Library

Training example

A library that is made for community. NeuralNetworkExample_FishersIris and NeuralNetworkExample_ImageRecognition are examples of how to use the library, NeuralNetwork is the project of the library and you can find the .dll in the folder "DLL is HERE". There you can find 2 folders with dll files for .Net Core and .Net Framework.

NeuralNetworkExample_FishersIris is the official example of usage of the NeuralNetwork, so I recommend using this project to test the performance of the Neural Network.

Great thanks to diademoff for the help with this project and for creating one of the examples of the usage of the Neural Network (NeuralNetworkExample_ImageRecognition).

Table of contents


Optimizations that has been done

  • Structs are more preferable than classes and objects.

  • All the variables are introduced in the body of the class, so garbage collector will execute less often.

  • A reference is used for passing structs in the arguments, therefore the program does not copy the struct every time the function is called.

  • For loop is more preferable than foreach loop. Also you can unroll loop.

  • Sometimes serial code is more preferable than functions.


Setting up the library

If you are using .Net Framework you must uncheck Prefer 32-bit in project preferences. Uncheck #1 or Uncheck #2

Add the DLL to your project (how to add dll in VS) and type this:

using NN;

How to use the library

Neuron layer types

There are 3 types of neuron layers:

  • Input layer
  • Hidden layer
  • Output layer

There may be one bias in each layer (except output). It's value is always '1' in every scenario.

NeuralNetwork creation

Firstly, you have to declare the struct of the Neural Network.

const string NeuronsAndLayers = "4+ 15+ 6 3"; 
// 4+  — input neurons and 1 bias neuron ([0]-InputNeurons)
// 15+ — hidden neurons in 1-st hidden layer and 1 bias neuron ([1]-Neurons In 1-st HiddenLayer) ([..])
// 6  — neurons in 2-nd hidden layer ([n-1]-Neurons In 2-nd HiddenLayer)
// 3  — output neurons ([n]-OutputNeurons)
// put + in each layer (except OUTPUT) to add bias

You must declare at least 2 layers of the Neural Network. [0] (Input) and [n] (Output).

  • To create a Neural Network you need to create an object of a class "NN.NeuralNetwork" using its constructor:

    NeuralNetwork network = new NeuralNetwork(NeuronsAndLayers, -1, 1);

    The first argument is the struct of the Neural Network (read above), after it you need to declare the bounds of random weights of synapses.

  • To change the Moment and LearningRate of the Neural Network:

    NeuralNetwork network = new NeuralNetwork(NeuronsAndLayers, -1, 1)
    {
        Moment = MyMoment, // from 0 to 1, for example 0.5
        LearningRate = MyLearningRate // from 0.1 to ∞, for example 1.2
    };

NeuralNetwork usage

  • To run the Neural Network use:

    Neuron[] endValue = network.RunNetwork(trainingData[i].IN);

    [Neuron[]](#neuron-struct) is a struct of a NeuralNetwork class. Each neuron has it's current value and outgoing synapse weights. The only argument of the function RunNetwork() is an array with Input values of Neurons. In this case, it may be [1, 1, 0, 1] (as it was written earlier, the current Neural Network has only 4 Input Neurons). The function returns Output value of Neurons, so you can interpret it as you wish

  • To teach the Network by current example (to make one iteration of teaching by current training example) use this function:

network.TeachNetwork(trainingData[i].OUT);

Where trainingData[i].OUT is an array with expected output value of neurons. In this case, it may be [0, 1, 0] (as it was written earlier, the current Neural Network has only 3 output neurons).


NeuralNetwork Load/Save

  • To Load the Neural Network use this constructor:

    NeuralNetwork network = new NeuralNetwork(@"C:\s\MyNeural.aaa");

    The only argument of this function is a path to the Neural Network file (as you could guess).

  • To Save the Neural Network use this method:

    network.SaveNetwork(@"C:\s\Neural.aaa");

    The only argument of this function is a path where the Network will be saved and the name of the Neural Network file (as you could guess).


Example of each teach iteration

  • Creating a new Neural Network (or you can load it if you already have one):

    string NeuronsAndLayers = "7 8+ 6 5 3";
    NeuralNetwork network = new NeuralNetwork(NeuronsAndLayers, -1, 1)
    {
      Moment = 0.5,
      LearningRate = 0.7
    };
  • Declaring input data:

    double[] inputData = new double[] { 1, 0, 1, 1, 0, 1, 1 };
  • Declaring output data:

    double[] outputData = new double[] { 1, 0, 1 };
  • Neural network teaching:

    Neuron[] outputNeurons = network.RunNetwork(inputData);
    network.TeachNetwork(outputData);
  • Alternative teaching (so you don't have to Run the Network manually before Teaching):

    network.TeachNetwork(inputData, outputData);
  • Neural network running:

    Neuron[] outputNeurons = network.RunNetwork(inputData);
  • Counting the error of the outpt

    Neuron[] outputNeurons = network.RunNetwork(inputData);
    //Counting an error of current unit
    end = 0;
    for (uint neuronIndex = 0; neuronIndex < outputNeurons.Length; ++neuronIndex)
             end += Math.Pow(outputData[neuronIndex] - outputNeurons[neuronIndex].value, 2);
    error = end / outputData.Length; //((i1[expected]-a1[output])*(i1-a1)+...+(in-an)*(in-an))/n
    errorSum += error;

Do it in a loop for greater efficiency. For more information check the examples.


Contributing

You can freely use the library if you want.

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.


Notes

Neuron struct

[Serializable]
public struct Neuron
{
      public double value;
      public double[] weights;
}