用java写bp神经网络（四）

接上篇。

在(一)和(二)中，程序的体系是Net,Propagation,Trainer,Learner,DataProvider。这篇重构这个体系。

Net

首先是Net，在上篇重新定义了激活函数和误差函数后，内容大致是这样的：

List<DoubleMatrix> weights = new ArrayList<DoubleMatrix>();
	List<DoubleMatrix> bs = new ArrayList<>();
	List<ActivationFunction> activations = new ArrayList<>();
	CostFunction costFunc;
	CostFunction accuracyFunc;
	int[] nodesNum;
	int layersNum;

public CompactDoubleMatrix getCompact(){
		return new CompactDoubleMatrix(this.weights,this.bs);
	}

函数getCompact()生成对应的超矩阵。

DataProvider

DataProvider是数据的提供者。

public interface DataProvider {
    DoubleMatrix getInput();
    DoubleMatrix getTarget();
}

如果输入为向量，还包含一个向量字典。

public interface DictDataProvider extends DataProvider {
	public DoubleMatrix getIndexs();
	public DoubleMatrix getDict();
}

每一列为一个样本。getIndexs()返回输入向量在字典中的索引。

我写了一个有用的类BatchDataProviderFactory来对样本进行批量分割，分割成minibatch。

int batchSize;
	int dataLen;
	DataProvider originalProvider;
	List<Integer> endPositions;
	List<DataProvider> providers;

	public BatchDataProviderFactory(int batchSize, DataProvider originalProvider) {
		super();
		this.batchSize = batchSize;
		this.originalProvider = originalProvider;
		this.dataLen = this.originalProvider.getTarget().columns;
		this.initEndPositions();
		this.initProviders();
	}

	public BatchDataProviderFactory(DataProvider originalProvider) {
		this(4, originalProvider);
	}

	public List<DataProvider> getProviders() {
		return providers;
	}

batchSize指明要分多少批，getProviders返回生成的minibatch，被分的原始数据为originalProvider。

Propagation

Propagation负责对神经网络的正向传播过程和反向传播过程。接口定义如下：

public interface Propagation {
	public PropagationResult propagate(Net net,DataProvider provider);
}

传播函数propagate用指定数据对指定网络进行传播操作，返回执行结果。

BasePropagation实现了该接口，实现了简单的反向传播：

public class BasePropagation implements Propagation{

	// 多个样本。
	protected ForwardResult forward(Net net,DoubleMatrix input) {

		ForwardResult result = new ForwardResult();
		result.input = input;
		DoubleMatrix currentResult = input;
		int index = -1;
		for (DoubleMatrix weight : net.weights) {
			index++;
			DoubleMatrix b = net.bs.get(index);
			final ActivationFunction activation = net.activations
					.get(index);
			currentResult = weight.mmul(currentResult).addColumnVector(b);
			result.netResult.add(currentResult);

			// 乘以导数
			DoubleMatrix derivative = activation.derivativeAt(currentResult);
			result.derivativeResult.add(derivative);

			currentResult = activation.valueAt(currentResult);
			result.finalResult.add(currentResult);

		}

		result.netResult=null;// 不再需要。

		return result;
	}

    // 多个样本梯度平均值。
	protected BackwardResult backward(Net net,DoubleMatrix target,
			ForwardResult forwardResult) {
		BackwardResult result = new BackwardResult();

		DoubleMatrix output = forwardResult.getOutput();
		DoubleMatrix outputDerivative = forwardResult.getOutputDerivative();

		result.cost = net.costFunc.valueAt(output, target);
		DoubleMatrix outputDelta = net.costFunc.derivativeAt(output, target).muli(outputDerivative);
		if (net.accuracyFunc != null) {
			result.accuracy=net.accuracyFunc.valueAt(output, target);
		}

		result.deltas.add(outputDelta);
		for (int i = net.layersNum - 1; i >= 0; i--) {
			DoubleMatrix pdelta = result.deltas.get(result.deltas.size() - 1);

			// 梯度计算，取所有样本平均
			DoubleMatrix layerInput = i == 0 ? forwardResult.input
					: forwardResult.finalResult.get(i - 1);
			DoubleMatrix gradient = pdelta.mmul(layerInput.transpose()).div(
					target.columns);
			result.gradients.add(gradient);
			// 偏置梯度
			result.biasGradients.add(pdelta.rowMeans());

			// 计算前一层delta，若i=0，delta为输入层误差，即input调整梯度，不作平均处理。
			DoubleMatrix delta = net.weights.get(i).transpose().mmul(pdelta);
			if (i > 0)
				delta = delta.muli(forwardResult.derivativeResult.get(i - 1));
			result.deltas.add(delta);
		}
		Collections.reverse(result.gradients);
		Collections.reverse(result.biasGradients);

		//其它的delta都不需要。
		DoubleMatrix inputDeltas=result.deltas.get(result.deltas.size()-1);
		result.deltas.clear();
		result.deltas.add(inputDeltas);

		return result;
	}

	@Override
	public PropagationResult propagate(Net net, DataProvider provider) {
		ForwardResult forwardResult=this.forward(net, provider.getInput());
		BackwardResult backwardResult=this.backward(net, provider.getTarget(), forwardResult);
		PropagationResult result=new PropagationResult(backwardResult);
		result.output=forwardResult.getOutput();
		return result;
	}

我们定义的PropagationResult略为：

public class PropagationResult{
		DoubleMatrix output;// 输出结果矩阵：outputLen*sampleLength
		DoubleMatrix cost;// 误差矩阵：1*sampleLength
		DoubleMatrix accuracy;// 准确度矩阵：1*sampleLength
		private List<DoubleMatrix> gradients;// 权重梯度矩阵
		private List<DoubleMatrix> biasGradients;// 偏置梯度矩阵
		DoubleMatrix inputDeltas;//输入层delta矩阵：inputLen*sampleLength

		public CompactDoubleMatrix getCompact(){
			return new CompactDoubleMatrix(gradients,biasGradients);
		}

	}

另一个实现了该接口的类为MiniBatchPropagation。他在内部用并行方式对样本进行传播，然后对每个minipatch结果进行综合，内部用到了BatchDataProviderFactory类和BasePropagation类。

Trainer

Trainer接口定义为：

public interface Trainer {
    public void train(Net net,DataProvider provider);
}

简单的实现类为：

public class CommonTrainer implements Trainer {
	int ecophs;
	Learner learner;
	Propagation propagation;
	List<Double> costs = new ArrayList<>();
	List<Double> accuracys = new ArrayList<>();
	public void trainOne(Net net, DataProvider provider) {
		PropagationResult propResult = this.propagation
				.propagate(net, provider);
		learner.learn(net, propResult, provider);

		Double cost = propResult.getMeanCost();
		Double accuracy = propResult.getMeanAccuracy();
		if (cost != null)
			costs.add(cost);
		if (accuracy != null)
			accuracys.add(accuracy);
	}

	@Override
	public void train(Net net, DataProvider provider) {
		for (int i = 0; i < this.ecophs; i++) {
			System.out.println("echops:"+i);
			this.trainOne(net, provider);
		}

	}
}

简单的迭代echops此，没有智能停止功能，每次迭代用Learner调节权重。

Learner

Learner根据每次传播结果对网络权重进行调整，接口定义如下：

public interface Learner<N extends Net,P extends DataProvider> {
    public void learn(N net,PropagationResult propResult,P provider);
}

一个简单的根据动量因子-自适应学习率进行调整的实现类为：

public class MomentAdaptLearner<N extends Net, P extends DataProvider>
		implements Learner<N, P> {
	double moment = 0.7;
	double lmd = 1.05;
	double preCost = 0;
	double eta = 0.01;
	double currentEta = eta;
	double currentMoment = moment;
	CompactDoubleMatrix preGradient;

	public MomentAdaptLearner(double moment, double eta) {
		super();
		this.moment = moment;
		this.eta = eta;
		this.currentEta = eta;
		this.currentMoment = moment;
	}

	public MomentAdaptLearner() {

	}

	@Override
	public void learn(N net, PropagationResult propResult, P provider) {
		if (this.preGradient == null)
			init(net, propResult, provider);

		double cost = propResult.getMeanCost();
		this.modifyParameter(cost);
		System.out.println("current eta:" + this.currentEta);
		System.out.println("current moment:" + this.currentMoment);
		this.updateGradient(net, propResult, provider);

	}

	public void updateGradient(N net, PropagationResult propResult, P provider) {
		CompactDoubleMatrix netCompact = this.getNetCompact(net, propResult,
				provider);
		CompactDoubleMatrix gradCompact = this.getGradientCompact(net,
				propResult, provider);
		gradCompact = gradCompact.mul(currentEta * (1 - currentMoment)).addi(
				preGradient.mul(currentMoment));
		netCompact.subi(gradCompact);
		this.preGradient = gradCompact;
	}

	public CompactDoubleMatrix getNetCompact(N net,
			PropagationResult propResult, P provider) {
		return net.getCompact();
	}

	public CompactDoubleMatrix getGradientCompact(N net,
			PropagationResult propResult, P provider) {
		return propResult.getCompact();
	}

	public void modifyParameter(double cost) {

		if (this.currentEta > 10) {
			this.currentEta = 10;
		} else if (this.currentEta < 0.0001) {
			this.currentEta = 0.0001;
		} else if (cost < this.preCost) {
			this.currentEta *= 1.05;
			this.currentMoment = moment;
		} else if (cost < 1.04 * this.preCost) {
			this.currentEta *= 0.7;
			this.currentMoment *= 0.7;
		} else {
			this.currentEta = eta;
			this.currentMoment = 0.1;
		}
		this.preCost = cost;
	}

	public void init(Net net, PropagationResult propResult, P provider) {
		PropagationResult pResult = new PropagationResult(net);
		preGradient = pResult.getCompact().dup();
	}

}

在上面的代码中，我们可以看到CompactDoubleMatrix类对权重自变量的封装，使代码更加简洁，它在此表现出来的就是一个超矩阵，超向量，完全忽略了内部的结构。

同时，其子类实现了同步更新字典的功能，代码也很简洁，只是简单的把需要调整的矩阵append到超矩阵中去即可，在父类中会统一对其进行调整：

public class DictMomentLearner extends
		MomentAdaptLearner<Net, DictDataProvider> {

	public DictMomentLearner(double moment, double eta) {
		super(moment, eta);
	}

	public DictMomentLearner() {
		super();
	}

	@Override
	public CompactDoubleMatrix getNetCompact(Net net,
			PropagationResult propResult, DictDataProvider provider) {
		CompactDoubleMatrix result = super.getNetCompact(net, propResult,
				provider);
		result.append(provider.getDict());
		return result;
	}

	@Override
	public CompactDoubleMatrix getGradientCompact(Net net,
			PropagationResult propResult, DictDataProvider provider) {
		CompactDoubleMatrix result = super.getGradientCompact(net, propResult,
				provider);
		result.append(DictUtil.getDictGradient(provider, propResult));
		return result;
	}

	@Override
	public void init(Net net, PropagationResult propResult,
			DictDataProvider provider) {
		DoubleMatrix preDictGradient = DoubleMatrix.zeros(
				provider.getDict().rows, provider.getDict().columns);
		super.init(net, propResult, provider);
		this.preGradient.append(preDictGradient);
	}
}