NeuralNetworks.Unity/Assets/Perceptron.cs

using System;
using System.Collections;
using System.Collections.Generic;
using UnityEngine;

public class Perceptron : MonoBehaviour
{
    public List<Layer> Layers;
    public List<Synapses> _synapses;

    [System.Serializable]
    public class Layer {
        public int Size;
        public bool AddOffsetNeuron;
        public int TotalSize {
            get {
                return AddOffsetNeuron ? (Size + 1) : Size;
            }
        }
        public void Init() {
            Values = new float[TotalSize];
            Errors = new float[TotalSize];


            if (AddOffsetNeuron) {
                Values[Values.Length - 1] = 1.0f;
            }
        }

        public float[] Values;
        public float[] Errors;
    }



    [System.Serializable]
    public class Synapses {
        public float[,] Values;
        public int InSize;
        public int OutSize;

        public Synapses(int inSize, int outSize) {
            InSize = inSize;
            OutSize = outSize;
            Values = new float[inSize, outSize];
            SetRandom();
          //  SetOne();
            Debug.Log($"Make synapses: {inSize}x{outSize}");
        }

        public void SetOne() {

            for (int i = 0; i < InSize; ++i) {
                for (int j = 0; j < OutSize; ++j) {
                    Values[i, j] = 1;
                }
            }
        }

        public void SetRandom() {

            for(int i = 0; i < InSize; ++i) {
                for (int j = 0; j < OutSize; ++j) {
                    Values[i, j] = UnityEngine.Random.value * 2.0f - 1.0f;
                }
            }
        } 
    }

    const float FxCapDerivative = 0.01f;
    //float Fx(float x) {
    //    if(x > 1.0f) {
    //        return 1.0f + FxCapDerivative * (x - 1.0f);
    //    }else if(x < 0.0f) {
    //        return FxCapDerivative * x;
    //    }
    //    return x;
    //}
    //
    //float FxDerivative(float x) {
    //    if (x > 1.0f) {
    //        return FxCapDerivative;
    //    } else if (x < 0.0f) {
    //        return FxCapDerivative;
    //    }
    //    return 1.0f;
    //}

    float Fx(float x) {
        return 1.0f / (1.0f + Mathf.Exp(-x));
    }

    float FxDerivative(float x) {
        return x * (1.0f - x);
    }



    public void PropagateErrors() {
        for (int inLayerId = Layers.Count - 2; inLayerId >= 0; --inLayerId) {
            int outLayerId = inLayerId + 1;

            var inLayer = Layers[inLayerId];
            var outLayer = Layers[outLayerId];
            var weights = _synapses[inLayerId];

            for(int i = 0; i < inLayer.Size; ++i) {
                float err = 0.0f;
                for (int o = 0; o < outLayer.Size; ++o) {
                    err += outLayer.Errors[o] * weights.Values[i, o];
                }
                inLayer.Errors[i] = err;
            }
        }
    }

    public void RecalculateWeights(float learnRate) {
        for (int inLayerId = 0; inLayerId < Layers.Count - 1; ++inLayerId) {
            int outLayerId = inLayerId + 1;

            var inLayer = Layers[inLayerId];
            var outLayer = Layers[outLayerId];
            var weights = _synapses[inLayerId];
            for (int i = 0; i < inLayer.TotalSize; ++i) {
 
                for (int o = 0; o < outLayer.Size; ++o) {

                    weights.Values[i, o] += (learnRate * outLayer.Errors[o]) * FxDerivative(outLayer.Values[o]) * inLayer.Values[i];
                }
            }
        }
    }

    public void CalculateOutputs() {
        for (int inLayerId = 0; inLayerId < Layers.Count - 1; ++inLayerId) {
            int outLayerId = inLayerId + 1;

            var inLayer = Layers[inLayerId];
            var outLayer = Layers[outLayerId];
            var weights = _synapses[inLayerId];
            for (int o = 0; o < outLayer.Size; ++o) {

                float sum = 0;
                for (int i = 0; i < inLayer.TotalSize; ++i) {
                    sum += weights.Values[i, o] * inLayer.Values[i];
                }
                outLayer.Values[o] = Fx(sum);
            }
        }
    }


    public void Learn(float learnRate) {
        PropagateErrors();
        RecalculateWeights(learnRate);
    }

    void Start()
    {
        foreach(var layer in Layers) {
            layer.Init();
        }

        _synapses = new List<Synapses>();
        for (int i = 0; i < Layers.Count - 1; ++i) {
            _synapses.Add(new Synapses(Layers[i].TotalSize, Layers[i + 1].Size));
        }
    }

    void Update()
    {
        
    }
}