private List <Mat <double> > ApplyPooling(List <Mat <double> > inputImages)
{
List <Mat <double> > pooledImages = new List <Mat <double> >();
foreach (Mat <double> inputImage in inputImages)
{
if ((inputImage.Row & 1) != 0 || (inputImage.Column & 1) != 0)
{
throw new Exception();
}
int poolHeight = inputImage.Row >> 1;
int poolWidth = inputImage.Column >> 1;
double[] mask = new double[4];
Mat <double> pooledImage = new Mat <double>(poolHeight, poolWidth);
for (int y = 0; y < inputImage.Row - 1; y++)
{
for (int x = 0; x < inputImage.Column - 1; x++)
{
for (int dy = 0; dy < 2; dy++)
{
for (int dx = 0; dx < 2; dx++)
{
mask[dy * 2 + dx] = inputImage.Element[y + dy, x + dx];
}
}
Array.Sort(mask);
pooledImage.Element[y / 2, x / 2] = mask[3];
}
}
pooledImages.Add(pooledImage);
}
return(pooledImages);
}
public double Train(Mat <double> image, double[] targetData)
{
if ((image.Row != imageSize.height) ||
(image.Column != imageSize.width))
{
throw new Exception();
}
int convLevel = 0;
int outputLevel = 0;
List <Mat <double> > inputImages = new List <Mat <double> >();
inputImages.Add(image);
convLayersOutput.Clear();
convLayersOutput.Add(inputImages);
foreach (ECNNLayer layer in layers)
{
switch (layer)
{
case ECNNLayer.ConvLayer:
inputImages = ApplyConvolution(inputImages, kernels[convLevel], strides[convLevel]);
inputImages = ApplyActivations(inputImages, EActivation.ReLU);
convLayersOutput.Add(inputImages);
convLevel++;
break;
case ECNNLayer.PoolingLayer:
inputImages = ApplyPooling(inputImages);
convLayersOutput.Add(inputImages);
break;
}
outputLevel++;
}
fcLayersOutput.Clear();
int inputNodesCount = layersNodes[0];
double[] inputNodes = new double[inputNodesCount];
foreach (Mat <double> inputImage in inputImages)
{
double[] arr = Mat <double> .ConvertToArr(inputImage);
Array.Copy(arr, inputNodes, arr.Length);
}
Mat <double> nodeMat = new Mat <double>(1, inputNodesCount, inputNodes);
fcLayersOutput.Add(nodeMat);
int layerDepth = 0;
double ratio = 0.85;
for (; layerDepth < fcDepth - 2; layerDepth++)
{
Mat <double> netMat = Mat <double> .Mul(fcLayersOutput[layerDepth], nodeWeights[layerDepth]);
AddBias(netMat, biasWeights[layerDepth]);
Dropout(netMat, ratio);
ApplyActivation(netMat, EActivation.ReLU);
fcLayersOutput.Add(netMat);
}
Mat <double> outMat = Mat <double> .Mul(fcLayersOutput[layerDepth], nodeWeights[layerDepth]);
AddBias(outMat, biasWeights[layerDepth]);
ApplyActivation(outMat, EActivation.Sigmoid);
fcLayersOutput.Add(outMat);
double[] output = Mat <double> .ConvertToArr(outMat);
double[] errors = MSE(output, targetData);
double errorSum = 0;
foreach (double val in errors)
{
errorSum += val;
}
for (int i = 0; i < fcDepth - 1; i++)
{
double[] outputMat = Mat <double> .ConvertToArr(fcLayersOutput[fcLayersOutput.Count - 1 - i]);
double[] inputMat = Mat <double> .ConvertToArr(fcLayersOutput[fcLayersOutput.Count - 1 - i - 1]);
errors = UpdateWeight(layersNodes[fcDepth - i - 2], layersNodes[fcDepth - i - 1],
errors,
outputMat,
inputMat,
nodeWeights[fcDepth - i - 2], biasWeights[fcDepth - i - 2],
i == 0 ? true : false);
}
List <List <Mat <double> > > convGradients = new List <List <Mat <double> > >();
List <Mat <double> > errorMats = Mat <double> .ConvertToMats(errors, kernelLen);
for (int i = convLayersOutput.Count - 1; i > 0; i--)
{
switch (layers[i - 1])
{
case ECNNLayer.ConvLayer:
// i 번째가 출력 , i - 1번째가 입력?
convGradients.Add(errorMats);
convLevel--;
UpdateKernel(errorMats, convLayersOutput[i], convLayersOutput[i - 1], convLevel);
errorMats = PropagateConvGradient(errorMats, convLayersOutput[i], convLevel);
break;
case ECNNLayer.PoolingLayer:
errorMats = PropagatePoolGradient(errorMats, convLayersOutput[i], convLayersOutput[i - 1]);
break;
}
}
return(errorSum);
}
public double[] Query(Mat <double> image)
{
if ((image.Row != imageSize.height) ||
(image.Column != imageSize.width))
{
throw new Exception();
}
int convLevel = 0;
int outputLevel = 0;
List <Mat <double> > inputImages = new List <Mat <double> >();
inputImages.Add(image);
foreach (ECNNLayer layer in layers)
{
switch (layer)
{
case ECNNLayer.ConvLayer:
inputImages = ApplyConvolution(inputImages, kernels[convLevel], strides[convLevel]);
inputImages = ApplyActivations(inputImages, EActivation.ReLU);
convLevel++;
break;
case ECNNLayer.PoolingLayer:
inputImages = ApplyPooling(inputImages);
break;
}
outputLevel++;
}
fcLayersOutput.Clear();
int inputNodesCount = layersNodes[0];
double[] inputNodes = new double[inputNodesCount];
foreach (Mat <double> inputImage in inputImages)
{
double[] arr = Mat <double> .ConvertToArr(inputImage);
Array.Copy(arr, inputNodes, arr.Length);
}
Mat <double> nodeMat = new Mat <double>(1, inputNodesCount, inputNodes);
fcLayersOutput.Add(nodeMat);
int layerDepth = 0;
for (; layerDepth < fcDepth - 2; layerDepth++)
{
Mat <double> netMat = Mat <double> .Mul(fcLayersOutput[layerDepth], nodeWeights[layerDepth]);
AddBias(netMat, biasWeights[layerDepth]);
ApplyActivation(netMat, EActivation.ReLU);
fcLayersOutput.Add(netMat);
}
Mat <double> outMat = Mat <double> .Mul(fcLayersOutput[layerDepth], nodeWeights[layerDepth]);
AddBias(outMat, biasWeights[layerDepth]);
ApplyActivation(outMat, EActivation.Sigmoid);
fcLayersOutput.Add(outMat);
double[] output = Mat <double> .ConvertToArr(outMat);
return(output);
}
