public LConvArry(LayerParamsInfo paramInfo, int sampleCount)
{
ParamsInfo = paramInfo;
CoreArry = new LConvolution[sampleCount];
for (int index = 0; index < sampleCount; index++)
{
CoreArry[index] = new LConvolution(paramInfo);
}
}
/// <summary>
///
/// </summary>
/// <param name="layer">准备使用隐藏层数量</param>
/// <param name="count"></param>
/// <param name="depth"></param>
public void InitWeight(int layer, int[] count, int[] depth)
{
LayerInfo = new LayerParamsInfo[layer];
for (int l = 0; l < layer; l++)
{
LayerInfo[l] = new LayerParamsInfo();
LayerInfo[l].CurrentLayer = l;
LayerInfo[l].InitWeightBias(count[l], depth[l]);
}
}
public LConvolution(LayerParamsInfo parInfo)
{
ParamsInfo = parInfo;
WehtInfo = parInfo.WehtInfo;
this.kernelCount = WehtInfo.Length;
this.kernelDepth = WehtInfo[0].Length;
this.PadRow = WehtInfo[0][0].W.GetLength(0);
this.PadColumn = WehtInfo[0][0].W.GetLength(1);
this.Padding = 0;
this.Stride = 1;
kernelRow = WehtInfo[0][0].W.GetLength(0);
kernelColumn = WehtInfo[0][0].W.GetLength(1);
//卷积核的深度和单个样本的通道数是一致的。
this._col = new LMatrix[kernelDepth];
}
/// <summary>
/// 当前层数据进行卷积、激活、池化操作
/// </summary>
/// <param name="input">输入数据</param>
/// <param name="layerIndex">使用第几层卷积数据</param>
/// <returns></returns>
public double[,,,] forward(double[,,,] input, int layerIndex)
{
//Stopwatch sw = new Stopwatch();
//sw.Start();
//第一层卷积
LayerParamsInfo curLPInfo = LayerInfo[layerIndex];
//int kernelCount = curLayerInfo.Weights.Length;
int sampleCount = input.GetLength(0);
int kernelDepth = curLPInfo.WehtInfo.Length;
double[,,,] oneConvValues = null;
if (ConvLayer[layerIndex] == null)
{
ConvLayer[layerIndex] = new LConvArry(curLPInfo, sampleCount);
}
oneConvValues = ConvLayer[layerIndex].forward(input);
////激活层
//if (ActivLayer[layerIndex] == null)
// ActivLayer[layerIndex] = new LRelu[sampleCount][][];
//double[][][] oneActivValues = new double[kernelDepth][][];
//for (int index = 0; index < sampleCount; index++)
//{
// if (ActivLayer[layerIndex][index] == null)
// ActivLayer[layerIndex][index] = new LRelu[kernelDepth][][][];
// oneActivValues[index] = ActivLayer[layerIndex][index].forward(oneConvValues[i], false).Matrix;
//}
double[,,,] oneActivValues = null;
if (ActiveLayer[layerIndex] == null)
{
ActiveLayer[layerIndex] = new LActiveArry();
}
oneActivValues = ActiveLayer[layerIndex].forward(oneConvValues);
//池化层
double[,,,] onePoolValues = null;
if (PoolLayer[layerIndex] == null)
{
PoolLayer[layerIndex] = new LMaxPool();
}
onePoolValues = PoolLayer[layerIndex].forward(oneActivValues);
//sw.Stop();
//Debug.Write("正向传播："+sw.Elapsed.TotalMilliseconds.ToString() + "\n");
return(onePoolValues);
}
public void StartTrain(int epoch)
{
Sources = new CnnSource(epoch);
Sources.CollectImgInfo(@"D:\AI\test_img", @"D:\AI\test_img2");
Tuple <double[, , , ], double[, ]> Train = Sources.GetNextEpoch();
MapInput = Train.Item1;
LMatrix lable = Train.Item2;
#region 初始化隐藏层基本属性和隐藏层权重值
var layers = new LayerAttribute[] {
new LayerAttribute
{
Count = 26,
Depth = 3
},
new LayerAttribute
{
Count = 52,
Depth = 26
},
new LayerAttribute
{
Count = 52,
Depth = 38
}
};
LayerInfo = new LayerParamsInfo[layers.Length];
for (int index = 0; index < layers.Length; index++)
{
LayerInfo[index] = new LayerParamsInfo();
LayerInfo[index].CurrentLayer = index;
LayerInfo[index].InitWeightBias(layers[index].Count, layers[index].Depth);
}
#endregion
#region 准备训练对象
LOptimizer optimizer = new LAdam(0.1);
ThreeLayerNet net = null;
bool isRead = false;
string path = @"d:\wen\good\";
for (int i = 0; i < 50000; i++)
{
Debug.Write("开始时间" + DateTime.Now.ToString("HH:mm:ss\n"));
var inputFeature = neuralnetworkforward();
if (net == null)
{
net = new ThreeLayerNet(inputFeature.GetLength(1), 200, 2, isRead: isRead);
}
net.Input = inputFeature;
double loss = net.forward(lable);
Debug.Write(string.Format("{0}:{1}当前误差：{2}\n", DateTime.Now.ToString("HH:mm:ss"), i, loss));
double accuracy = net.Accuracy(lable);
Debug.Write(string.Format("{0}:{1}当前识别精度：{2}\n", DateTime.Now.ToString("HH:mm:ss"), i, accuracy));
LMatrix dout = net.backward(lable);
List <LMatrix> grads = net.Gradient();
List <LMatrix> param = net.GetParams();
string strWB = JsonConvert.SerializeObject(param.Select(x => x.Matrix));
if (!Directory.Exists(path + i.ToString()))
{
Directory.CreateDirectory(path + i.ToString());
}
File.WriteAllText(path + i.ToString() + @"\wb_" + accuracy.ToString() + "_.txt", strWB);
var _backdout = dout.ResetSize(38, CurRow, CurColumn);
neuralnetworkbackward(_backdout);
#region 卷积梯度更新
int infolen = LayerInfo.Length;
for (int idx = 0; idx < infolen; idx++)
{
LayerInfo[idx].UpdateGradient(param, grads);
}
#endregion
optimizer.Update(param, grads);
//isRead = true;
//记录每次的卷积值信息
File.WriteAllText(path + i.ToString() + @"\LayerInfo.txt", JsonConvert.SerializeObject(LayerInfo));
File.WriteAllText(path + i.ToString() + @"\ConvInfo.txt", JsonConvert.SerializeObject(ConvLayer));
}
#endregion
}
