using DenseCRF;
using FCN;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace CNN
{
 class Program
 {
 static void Main(string[] args)
 {
 

 float[,] X = new float[,] { { 0.5f}};
 float[,] Y = new float[,] { { -0.0053f } };
 float[,] W = new float[,] { { 0.1f } };
 // float[,] temp= Matrix.convnValid(X, W, 1, 0);
 float[,] temp = Matrix.convnValid(X, Y, 1, 0);

 float[,] TT = new float[,] { { -0.5160f, -0.5160f }, { -0.5160f, -0.7289f, } };
 float[,] Ta= MESBackwrad(X, Y, TT);
 
 temp = Matrix.convnValid(Ta, Y, 1, 0);
 // FCNTrain();
 FCNtest();
 // CNNCovTrain5x5();//全卷积训练
 //CNNFULLTrain5x5();//全连接训练
 // CNNtest();//全连接测试

 // CNNcovtest();//全卷积测试
 }
 static float[,] MESBackwrad(float[,] X, float[,] Y, float[,] temp)
 {
 int size = 2;
 float[,] W = new float[size, size];
 for (var i = 0; i < size; i++)
 for (var j = 0; j < size; j++)
 {
 float[,] ff = new float[temp.GetLength(0), temp.GetLength(1)];
 float sum = 0f;
 for (var x = i; x < i+size; x++)
 for (var y = j; y < j+size; y++)
 {
 //if (x >= Y.GetLength(0) || y >= Y.GetLength(1))
 // continue;
 ff[x-i, y-j] = -2 * X[x, y] * (Y[x - i, y - j] - temp[x - i, y - j]);
 sum += ff[x - i, y - j];
 }
 W[i, j] = sum / (size * 2);
 }

 return W;
 }
 static void FCNtest()
 {
 
 // outputlayer output = new outputlayer(false, 12, 2, 5);
 // output.isFull = true;
 // layers.Add(output);
 System.IO.StreamReader sr = new System.IO.StreamReader("tagjixue.fcn", Encoding.UTF8);
 string strmode = sr.ReadToEnd();
 sr.Close();
 cnnweights cnnw = Newtonsoft.Json.JsonConvert.DeserializeObject<cnnweights>(strmode);
 NeuralNetwork cnn = new NeuralNetwork();
 string file = "ii12.jpg";
 Matrix[] anno1 = util.readpnggetMatrix(file,false);
 int channl = anno1.GetLength(0);
 int wcount = 6;
 for (int ha = 0; ha < anno1.GetLength(0); ha++)
 {
 anno1[ha].values = ImgUtil.Bilinear(anno1[ha].values, 512, 512);
 
 }
 List<Layer> layers = new List<Layer>();

 layers.Add(new convlayer(1, 2, 5, 1, channl * wcount, false));//256
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//128
 layers.Add(new convlayer(1, 2, 5, channl * wcount, channl * wcount, false));//
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));// 64
 layers.Add(new convlayer(1, 2, 5, channl * wcount, channl * wcount, false));//
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//32
 layers.Add(new convlayer(1, 0, 5, channl * wcount, channl * wcount, false));//28
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//14
 layers.Add(new convlayer(1, 0, 5, channl * wcount, channl * wcount, false));//10
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//5 
 layers.Add(new outputlayer(false, channl * wcount, 2,5, false));
 cnn.FCNidentify(layers, cnnw, anno1, activateFun.sigmoid);

 }
 static void FCNTrain()
 {
 string[] Directs = System.IO.Directory.GetDirectories("res");

 int max=0;
 foreach (string dd in Directs)
 {
 string[] files = System.IO.Directory.GetFiles(dd); 
 int index = Convert.ToInt32(dd.Split('\\')[1]);
 if (max < index)
 max = index;


 }
 int channl = 3;
 List<Matrix> annolist = new List<Matrix>();
 List<lable> las = new List<lable>();
 foreach (string dd in Directs)
 {
 
 string[] files = System.IO.Directory.GetFiles(dd);
 int index = Convert.ToInt32(dd.Split('\\')[1]);
 
 

 List<Matrix> matrices = new List<Matrix>();
 foreach (string file in files)
 {
 lable la = new lable();
 la.tag = index;
 la.output = new float[max];
 la.output[index - 1] = 1;
 
 Matrix[] anno1 = util.readpnggetMatrix(file, false);
 channl = anno1.GetLength(0);
 for (int ha = 0; ha < anno1.GetLength(0); ha++)
 {
 anno1[ha].values = ImgUtil.Bilinear(anno1[ha].values, 256, 256);
 // im1[ha].values= ImgUtil.Bilinear(im1[ha].values, 64, 64);
 annolist.Add(anno1[ha]);
 las.Add(la);
 }
 
 


 }
 
 }
 List<Layer> layers = new List<Layer>();
 int wcount = 6;
 layers.Add(new convlayer(1, 2, 5, 1, channl* wcount));//256
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//128
 layers.Add(new convlayer(1, 2, 5, channl * wcount, channl * wcount));//
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));// 64
 layers.Add(new convlayer(1, 2, 5, channl * wcount, channl * wcount));//
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//32
 layers.Add(new convlayer(1, 0, 5, channl * wcount, channl * wcount));//28
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//14
 layers.Add(new convlayer(1, 0, 5, channl * wcount, channl * wcount));//10
 layers.Add(new Poolinglayer(2, channl * wcount, channl * wcount));//5 
 outputlayer output = new outputlayer(false, channl * wcount, max, 5);
 // output.isFull = true;
 layers.Add(output);
 NeuralNetwork cn = new NeuralNetwork();
 cnnweights cnnw = cn.Train(layers, annolist.ToArray(), las.ToArray(), 1,activateFun.sigmoid);
 string str = Newtonsoft.Json.JsonConvert.SerializeObject(cnnw);
 System.IO.StreamWriter sw = new System.IO.StreamWriter("tagjixue.fcn", false, Encoding.UTF8);
 sw.Write(str);
 sw.Close();
 }
 static void CNNcovtest()
 {
 Minst.MinstImgArr trainImg = Minst.read_Img("D:\\caffe\\Minst\\t10k-images.idx3-ubyte");
 Minst.MinstLabelArr trainLabel = Minst.read_Lable("D:\\caffe\\Minst\\t10k-labels.idx1-ubyte");
 List<Layer> layers = new List<Layer>();//28

 layers.Add(new convlayer(1, 0, 5, 1, 6));//24
 layers.Add(new Poolinglayer(2, 6, 6));//12
 layers.Add(new convlayer(1, 0, 5, 6, 12));//8
 layers.Add(new Poolinglayer(2, 12, 12));//4
 outputlayer output = new outputlayer(false, 12, 10,4);
 
 layers.Add(output);

 NeuralNetwork cnn = new NeuralNetwork();
 System.IO.StreamReader sr = new System.IO.StreamReader("tagjixue.mode", Encoding.UTF8);
 string strmode = sr.ReadToEnd();
 sr.Close();
 cnnweights cnnw = Newtonsoft.Json.JsonConvert.DeserializeObject<cnnweights>(strmode);
 Matrix anno = new Matrix();
 anno.values = trainImg.ImgPtr[1].ImgData;
 int max = cnn.CNNidentify(layers, cnnw, anno);
 }
 static void CNNtest()
 {
 Minst.MinstImgArr trainImg = Minst.read_Img("D:\\caffe\\Minst\\t10k-images.idx3-ubyte");
 Minst.MinstLabelArr trainLabel = Minst.read_Lable("D:\\caffe\\Minst\\t10k-labels.idx1-ubyte");
 List<Layer> layers = new List<Layer>();//28

 layers.Add(new convlayer( 1, 0, 5, 1, 6));//24
 layers.Add(new Poolinglayer( 2, 6, 6));//12
 layers.Add(new convlayer( 1, 0, 5, 6, 12));//8
 layers.Add(new Poolinglayer( 2, 12, 12));//4
 outputlayer output = new outputlayer(true,4 * 4 * 12, 10);
 output.isFull = true;
 layers.Add(output);

 NeuralNetwork cnn = new NeuralNetwork();
 System.IO.StreamReader sr = new System.IO.StreamReader("tagjixue.mode", Encoding.UTF8);
 string strmode = sr.ReadToEnd();
 sr.Close();
 cnnweights cnnw= Newtonsoft.Json.JsonConvert.DeserializeObject<cnnweights>(strmode);
 Matrix anno = new Matrix();
 anno.values = trainImg.ImgPtr[1].ImgData;
 int max=cnn.CNNidentify(layers, cnnw, anno);
 }
 static void CNNFULLTrain5x5()
 {
 Minst.MinstImgArr trainImg = Minst.read_Img("D:\\caffe\\Minst\\train-images.idx3-ubyte");
 Minst.MinstLabelArr trainLabel = Minst.read_Lable("D:\\caffe\\Minst\\train-labels.idx1-ubyte");


 List<Layer> layers = new List<Layer>();

 layers.Add(new convlayer(1, 0, 5, 1, 6));//32
 layers.Add(new Poolinglayer( 2, 6, 6));//16
 layers.Add(new convlayer(1, 0, 5, 6, 12));//16
 layers.Add(new Poolinglayer( 2, 12, 12));//14 
 outputlayer output = new outputlayer(true, 4 * 4 * 12, 10);
 // output.isFull = true;
 layers.Add(output);
 Matrix[] anno = new Matrix[50];
 lable[] lables = new lable[50];

 NeuralNetwork cn = new NeuralNetwork();
 for (int a = 0; a < 50; a++)
 {
 anno[a] = new Matrix();
 anno[a].values = trainImg.ImgPtr[a].ImgData;
 lables[a] = new lable();
 lables[a].output = trainLabel.LabelPtr[a].LabelData;
 lables[a].tag = trainLabel.LabelPtr[a].l;

 }
 cnnweights cnnw= cn.Train(layers, anno, lables, 1);
 string str = Newtonsoft.Json.JsonConvert.SerializeObject(cnnw);
 System.IO.StreamWriter sw = new System.IO.StreamWriter("tagjixue.mode", false, Encoding.UTF8);
 sw.Write(str);
 sw.Close();
 }
 static void CNNCovTrain5x5()
 {
 Minst.MinstImgArr trainImg = Minst.read_Img("D:\\caffe\\Minst\\train-images.idx3-ubyte");
 Minst.MinstLabelArr trainLabel = Minst.read_Lable("D:\\caffe\\Minst\\train-labels.idx1-ubyte");


 List<Layer> layers = new List<Layer>();

 layers.Add(new convlayer(1, 0, 5, 1, 6));//32
 layers.Add(new Poolinglayer(2, 6, 6));//16
 layers.Add(new convlayer(1, 0, 5, 6, 12));//16
 layers.Add(new Poolinglayer(2, 12, 12));//14 
 outputlayer output = new outputlayer(false, 12, 10,4);
 // output.isFull = true;
 layers.Add(output);
 Matrix[] anno = new Matrix[50];
 lable[] lables = new lable[50];

 NeuralNetwork cn = new NeuralNetwork();
 for (int a = 0; a < 50; a++)
 {
 anno[a] = new Matrix();
 anno[a].values = trainImg.ImgPtr[a].ImgData;
 lables[a] = new lable();
 lables[a].output = trainLabel.LabelPtr[a].LabelData;
 lables[a].tag = trainLabel.LabelPtr[a].l;

 }
 cnnweights cnnw = cn.Train(layers, anno, lables, 1);
 string str = Newtonsoft.Json.JsonConvert.SerializeObject(cnnw);
 System.IO.StreamWriter sw = new System.IO.StreamWriter("tagjixue.mode", false, Encoding.UTF8);
 sw.Write(str);
 sw.Close();
 }

 }
}