忍者ブログ

Memeplexes

プログラミング、3DCGとその他いろいろについて

かんたん!制限付きボルツマンマシン実装 実数バージョン (C#)訂正 [Deep Learningシリーズ]

可視層の状態をランダムに

以前、このような記事を書いたのですが、こっちのほうが正しいような気がしてきました。
しかし性能はどちらがいいかは…。

思い出しっぷりはこちらのほうが悪いですが、可視層もランダムにしようと思ったらこうなるでしょうね。

ちなみにこいつで画像を扱うと、結果がざらつくようです。
あんまり使いたくないですね。

RestrictedBoltzmannMachine.cs

using System;
using System.Threading.Tasks;

namespace RestrictedBoltzmannMachines.RealValue
{
    public class RestrictedBoltzmannMachine
    {
        public SymmetricConnection[][] Connections;
        public VisibleNeuron[] VisibleNeurons;
        public HiddenNeuron[] HiddenNeurons;

        public RestrictedBoltzmannMachine(int visibleNeuronCount, int hiddenNeuronCount, Random random) :
            this(SymmetricConnection.CreateRandomWeights(random, visibleNeuronCount, hiddenNeuronCount), new double[visibleNeuronCount], new double[hiddenNeuronCount], random)
        {
        }

        public RestrictedBoltzmannMachine(double[][] weights, double[] visibleBiases, double[] hiddenBiases, Random random)
        {
            this.VisibleNeurons = Neuron.CreateNeurons<VisibleNeuron>(visibleBiases);
            this.HiddenNeurons = Neuron.CreateNeurons<HiddenNeuron>(hiddenBiases);
            this.Connections = SymmetricConnection.CreateConnections(weights, VisibleNeurons, HiddenNeurons);
            Neuron.WireConnections(this.Connections);

            foreach (var neuron in this.VisibleNeurons)
            {
                neuron.Random = new Random(random.Next());
            }

            foreach (var neuron in this.HiddenNeurons)
            {
                neuron.Random = new Random(random.Next());
            }
        }

        public void SetVisibleNeuronValues(double[] visibleValues)
        {
            for (int i = 0; i < this.VisibleNeurons.Length; i++)
            {
                this.VisibleNeurons[i].Value = visibleValues[i];
            }
        }

        public void LearnFromData(double learningRate, int freeAssociationStepCount = 1)
        {
            Wake(learningRate);
            Sleep(learningRate, freeAssociationStepCount);
            EndLearning();
        }

        public void Wake(double learningRate)
        {
            UpdateHiddenNeurons();
            learn(learningRate);
        }

        public void UpdateVisibleNeurons()
        {
            updateNeurons(this.VisibleNeurons);
        }

        public void UpdateHiddenNeurons()
        {
            updateNeurons(this.HiddenNeurons);
        }

        private void updateNeurons(Neuron[] neurons)
        {
            Parallel.ForEach(neurons, neuron => neuron.Update());
        }


        private void learn(double learningRate)
        {
            Parallel.ForEach(Connections, connectionRow =>
            {
                foreach (var connection in connectionRow)
                {
                    connection.Learn(learningRate);
                }
            });

            Parallel.ForEach(VisibleNeurons, neuron => neuron.Learn(learningRate));
            Parallel.ForEach(HiddenNeurons, neuron => neuron.Learn(learningRate));
        }

        public void Sleep(double learningRate, int freeAssociationStepCount)
        {
            doFreeAssociation(freeAssociationStepCount);
            learn(-learningRate);
        }

        //Gibbs sampling
        private void doFreeAssociation(int freeAssociationStepCount)
        {
            for (int step = 0; step < freeAssociationStepCount; step++)
            {
                UpdateVisibleNeurons();
                UpdateHiddenNeurons();
            }
        }

        public void EndLearning()
        {
            Parallel.ForEach(Connections, connectionRow =>
            {
                foreach (var connection in connectionRow)
                {
                    connection.EndLearning();
                }
            });

            Parallel.ForEach(VisibleNeurons, neuron => neuron.EndLearning());
            Parallel.ForEach(HiddenNeurons, neuron => neuron.EndLearning());
        }

        public void Associate()
        {
            UpdateHiddenNeurons();
            UpdateVisibleNeurons();
        }
    }
}

Neuron.cs

using System;
using System.Collections.Generic;

namespace RestrictedBoltzmannMachines.RealValue
{
    public abstract class Neuron
    {
        public double Value;
        public double Bias;
        public double DeltaBias;
        public List<Synapse> Synapses = new List<Synapse>();
        public Random Random;

        public abstract void Update();
        public abstract void Learn(double learningRate);

        public void EndLearning()
        {
            this.Bias += this.DeltaBias;
            this.DeltaBias = 0;
        }

        public static T[] CreateNeurons<T>(double[] biases)
            where T : Neuron, new() 
        {
            T[] result = new T[biases.Length];

            for (int i = 0; i < result.Length; i++)
            {
                result[i] = new T { Bias = biases[i] };
            }

            return result;
        }

        public static void WireConnections(SymmetricConnection[][] connections)
        {
            foreach (var connectionRow in connections)
            {
                foreach (var connection in connectionRow)
                {
                    Synapse hiddenConnection = new Synapse();
                    hiddenConnection.Connection = connection;
                    hiddenConnection.SourceNeuron = connection.VisibleNeuron;
                    connection.HiddenNeuron.Synapses.Add(hiddenConnection);

                    Synapse visibleConnection = new Synapse();
                    visibleConnection.Connection = connection;
                    visibleConnection.SourceNeuron = connection.HiddenNeuron;
                    connection.VisibleNeuron.Synapses.Add(visibleConnection);
                }
            }
        }

        protected double GetInputFromSourceNeurons()
        {
            double result = 0;

            for (int i = 0; i < Synapses.Count; i++)
            {
                var s = Synapses[i];
                result += s.Connection.Weight * s.SourceNeuron.Value;
            }

            return result;
        }

        protected static double Sigmoid(double x)
        {
            return 1.0 / (1.0 + Math.Exp(-x));
        }
    }

    public class VisibleNeuron : Neuron
    {
        public override void Update()
        {
            this.Value = Sigmoid(nextGaussian(Random) + GetInputFromSourceNeurons() + Bias);
        }

        private static double nextGaussian(Random random)
        {
            return Math.Sqrt(-2.0 * Math.Log(random.NextDouble())) 
                * Math.Sin(2.0 * Math.PI * random.NextDouble()); 
        }

        public override void Learn(double learningRate)
        {
            this.DeltaBias += learningRate * this.Value;
        }
    }

    public class HiddenNeuron : Neuron
    {
        public double Probability;

        public override void Update()
        {
            this.Probability = Sigmoid(GetInputFromSourceNeurons() + Bias);
            this.Value = nextBool(Random, this.Probability) ? 1 : 0;
        }

        private static bool nextBool(Random random, double rate)
        {
            if (rate < 0 || 1 < rate) return false;
            return random.NextDouble() < rate;
        }

        public override void Learn(double learningRate)
        {
            this.DeltaBias += learningRate * this.Probability;
        }
    }
}

前回のソースコードと合わせると結果はこうなります:

0.99    0.46    0.94    0.04    0.00
0.01    0.10    0.89    0.66    0.93

まあいいんじゃないでしょうか?

拍手[0回]

PR