private void button1_Click(object sender, EventArgs e)
{
//以下为数据准备部分
string trainDataPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "train-images.idx3-ubyte");
string trainLabelPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "train-labels.idx1-ubyte");
string testDataPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "t10k-images.idx3-ubyte");
string testLabelPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "t10k-labels.idx1-ubyte");
MNISTDataProcessing dataProcessing = new MNISTDataProcessing(new MNISTDataConverter(new MNISTClassifier()));
MNISTEvaluater evaluater = new MNISTEvaluater(new MNISTDataConverter(new MNISTClassifier()));
var inputDataSet = dataProcessing.GetData(trainDataPath, trainLabelPath, 60000);
var processDataSet = dataProcessing.GetProcessDataCollection(60000, inputDataSet);
ProcessData[] trainDataSet;
processDataSet = processDataSet.Concat(processDataSet.Select(r => dataProcessing.HorizontalReverseImage(r))).ToList();
List <ProcessData> testDataSet1;
List <ProcessData> testDataSet2;
//trainDataSet = processDataSet.ToArray();//训练集
ProcessData[] verifyDataSet;
(trainDataSet, verifyDataSet) = dataProcessing.DivideCollection(0.9, processDataSet);
//ILookup<double, ProcessData> dataDistribution = trainDataSet.ToLookup(r => MNISTDataConverter.GetFromOnehot(r.Label[0, 0].ToRowMajorArray()));
testDataSet1 = dataProcessing.GetProcessDataCollection(5000, dataProcessing.GetData(testDataPath, testLabelPath, 5000)); //测试集1
testDataSet2 = dataProcessing.GetProcessDataCollection(10000, dataProcessing.GetData(testDataPath, testLabelPath, 10000)); //测试集2
listBox1.Items.Add(DateTime.Now + " : DataPreparation Completed.");
////模型建立部分
//Model avilaFNN = new Model(new CrossEntropyLoss(), "MNIST_LeNet_v1",
// //new NormaliztionLayer(trainDataSet, NormalizationMode.Standardization),
// new ConvolutionalLayer(new ReLU(), false, new Tuple<int, int>(28, 28), 5, 1, 6, ConvolutionMode.Narrow, new Normal(0, 1)),
// new PoolingLayer(3, false, 6, new Tuple<int, int>(24, 24), PoolingMode.Max),
// new ConvolutionalLayer(new Direct(0.01), false, new Tuple<int, int>(8, 8), 5, 6, 16, ConvolutionMode.Narrow, new Normal(0, 1)),
// new PoolingLayer(2, false, 16, new Tuple<int, int>(4, 4), PoolingMode.Average),
// new ConvolutionalLayer(new Direct(0.01), false, new Tuple<int, int>(2, 2), 2, 16, 120, ConvolutionMode.Narrow, new Normal(0, 1)),
// new PoolingLayer(1, true, 120, new Tuple<int, int>(1, 1), PoolingMode.Average),
// new FullConnectLayer(new Direct(1), 120, 120, true, new Normal(0, 1)),
// new FullConnectLayer(new Sigmoid(), 120, 156, false, new Normal(0, 2)),
// new SoftMaxLayer(156, 10, new Normal(0, 2))
// );
Model avilaFNN = new Model(null, "1").LoadModel <Model>(@"C:\Users\liu_y\source\repos\CSahrp\NeuralNetwork-v0.3\NeuralNetwork-v0.1\bin\Debug\Models\MNIST_CNN_v6-0.868.model");
listBox1.Items.Add(DateTime.Now + " : ModelBuild Completed.");
avilaFNN.layers[5].Locked = false;
avilaFNN.layers[6].Locked = false;
double maxAcc = 0.868;
Trainer trainer2 = new Trainer(1000, false, 0, new ExponentialDelayOptimizer(1, 0.9, 0.1));
////Trainer trainer2 = new Trainer(50, false, 0, new TriangularCyclicOptimizer(1500,0.2,0.005,0.98));
//for (int i = 0; i <= 1; i++)
//{
// trainDataSet = trainDataSet.SelectPermutation().ToArray();
// var temp = (Model)trainer2.Train(avilaFNN, trainDataSet, verifyDataSet);
// //Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
// //Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
// Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
// var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
// if (acc >= maxAcc)
// {
// avilaFNN = temp;
// avilaFNN.SaveModel(accuracy: acc);
// maxAcc = acc;
// }
//}
//chart1.DataSource = trainer2.lossTable;
//chart1.Series.Add("111");
//chart1.Series[0].XValueMember = trainer2.lossTable.Columns[0].ColumnName;
//chart1.Series[1].YValueMembers = trainer2.lossTable.Columns[1].ColumnName;
//chart1.Series[0].ChartType = SeriesChartType.Line;
//trainer2 = new Trainer(1000, false, 0, new ExponentialDelayOptimizer(1, 0.9, 0.05));
//for (int i = 0; i <= 4; i++)
//{
// trainDataSet = trainDataSet.SelectPermutation().ToArray();
// var temp = (Model)trainer2.Train(avilaFNN, trainDataSet,verifyDataSet);
// //Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
// //Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
// Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
// var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
// if (acc >= maxAcc)
// {
// avilaFNN = temp;
// avilaFNN.SaveModel(accuracy: acc);
// maxAcc = acc;
// }
//}
//trainer2 = new Trainer(500, false, 0, new ExponentialDelayOptimizer(0.5, 0.95, 0.05));
//for (int i = 0; i <= 4; i++)
//{
// trainDataSet = trainDataSet.SelectPermutation().ToArray();
// var temp = (Model)trainer2.Train(avilaFNN, trainDataSet,verifyDataSet);
// //Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
// //Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
// Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
// var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
// if (acc >= maxAcc)
// {
// avilaFNN = temp;
// avilaFNN.SaveModel(accuracy: acc);
// maxAcc = acc;
// }
//}
//trainer2 = new Trainer(1000, false, 0, new AdaGradOptimizer(1));
//for (int i = 0; i <= 4; i++)
//{
// trainDataSet = trainDataSet.SelectPermutation().ToArray();
// var temp = (Model)trainer2.Train(avilaFNN, trainDataSet);
// //Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
// //Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
// Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
// var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
// avilaFNN = temp;
// if (acc >= maxAcc)
// {
// avilaFNN.SaveModel(accuracy: acc);
// maxAcc = acc;
// }
//}
//trainer2 = new Trainer(1000, false, 0, new ExponentialDelayOptimizer(0.3, 0.9, 0.05));
//for (int i = 0; i <= 9; i++)
//{
// trainDataSet = trainDataSet.SelectPermutation().ToArray();
// var temp = (Model)trainer2.Train(avilaFNN, trainDataSet);
// //Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
// //Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
// Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
// var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
// avilaFNN = temp;
// {
// avilaFNN.SaveModel(accuracy: acc);
// maxAcc = acc;
// }
//}
trainer2 = new Trainer(400, false, 0, new ExponentialDelayOptimizer(0.01, 0.9, 0.05));
for (int i = 0; i <= 9; i++)
{
trainDataSet = trainDataSet.SelectPermutation().ToArray();
var temp = (Model)trainer2.Train(avilaFNN, trainDataSet);
//Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
//Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
avilaFNN = temp;
if (acc >= maxAcc)
{
avilaFNN.SaveModel(accuracy: acc);
maxAcc = acc;
}
}
trainer2 = new Trainer(300, false, 0, new ExponentialDelayOptimizer(0.005, 0.95, 0.05));
for (int i = 0; i <= 19; i++)
{
trainDataSet = trainDataSet.SelectPermutation().ToArray();
var temp = (Model)trainer2.Train(avilaFNN, trainDataSet, verifyDataSet);
//Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
//Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
Console.WriteLine(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet2, temp, new AvilaClassifier())}");
var acc = evaluater.EvaluateInDetails(testDataSet2, temp);
avilaFNN = temp;
if (acc >= maxAcc)
{
avilaFNN.SaveModel(accuracy: acc);
maxAcc = acc;
}
}
trainDataSet = trainDataSet.SelectPermutation().ToArray();
Trainer trainer4 = new Trainer(100, false, 0, new ExponentialDelayOptimizer(0.1, 0.9, 0.04));
avilaFNN = (Model)trainer4.Train(avilaFNN, trainDataSet);
Console.WriteLine(DateTime.Now + $" : 21th Training Completed, Accuracy is:{evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier())}");
var accuracy = evaluater.EvaluateInDetails(testDataSet2, avilaFNN);
avilaFNN.SaveModel(accuracy: accuracy);
////结果评估部分
//evaluater.EvaluateInDetails(testDataSet1, avilaFNN);
//Console.WriteLine("======================");
//Console.WriteLine("-----Difficult Test-----");
//var accuracy=evaluater.EvaluateInDetails(testDataSet2, avilaFNN);
//avilaFNN.SaveModel(accuracy: accuracy);
}
private void button2_Click(object sender, EventArgs e)
{
//以下为数据准备部分
string trainDataPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "train-images.idx3-ubyte");
string trainLabelPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "train-labels.idx1-ubyte");
string testDataPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "t10k-images.idx3-ubyte");
string testLabelPath = System.IO.Path.Combine(Application.StartupPath, "MNIST", "t10k-labels.idx1-ubyte");
MNISTDataProcessing dataProcessing = new MNISTDataProcessing(new MNISTDataConverter(new MNISTClassifier()));
MNISTEvaluater evaluater = new MNISTEvaluater(new MNISTDataConverter(new MNISTClassifier()));
var inputDataSet = dataProcessing.GetData(trainDataPath, trainLabelPath, 60000);
var processDataSet = dataProcessing.GetProcessDataCollection(60000, inputDataSet);
ProcessData[] trainDataSet;
List <ProcessData> testDataSet1;
List <ProcessData> testDataSet2;
trainDataSet = processDataSet.ToArray();//训练集
ILookup <double, ProcessData> dataDistribution = trainDataSet.ToLookup(r => MNISTDataConverter.GetFromOnehot(r.Label[0, 0].ToRowMajorArray()));
testDataSet1 = dataProcessing.GetProcessDataCollection(5000, dataProcessing.GetData(testDataPath, testLabelPath, 5000)); //测试集1
testDataSet2 = dataProcessing.GetProcessDataCollection(10000, dataProcessing.GetData(testDataPath, testLabelPath, 10000)); //测试集2
listBox1.Items.Add(DateTime.Now + " : DataPreparation Completed.");
Model avilaFNN = new Model(null, "1").LoadModel <Model>(@"C:\Users\liu_y\source\repos\CSahrp\NeuralNetwork-v0.3\NeuralNetwork-v0.1\bin\Debug\Models\MNIST_CNN_v2-0.867.model");
listBox1.Items.Add(DateTime.Now + " : ModelLoad Completed.");
////模型建立部分
//Model avilaFNN = new Model(new CrossEntropyLoss(), "MNIST_CNN_v1",
// new ConvolutionalLayer(new Direct(), false, new Tuple<int, int>(28, 28), 5, 1, 3, ConvolutionMode.Narrow, new Normal(0, 2)),
// new PoolingLayer(2, false, 3, new Tuple<int, int>(24, 24), PoolingMode.Max),
// new ConvolutionalLayer(new Direct(), false, new Tuple<int, int>(12, 12), 5, 3, 5, ConvolutionMode.Narrow, new Normal(0, 2)),
// new PoolingLayer(2, true, 5, new Tuple<int, int>(8, 8), PoolingMode.Max),
// new FullConnectLayer(new Direct(), 80, 80, true, new Normal(0, 1)),
// new FullConnectLayer(new Sigmoid(), 80, 32, false, new Normal(0, 2)),
// new SoftMaxLayer(32, 10, new Normal(0.5, 1))
// );
//Trainer trainer1 = new Trainer(400, false, 0, new ExponentialDelayOptimizer(2, 0.9, 0.08));
//avilaFNN = (Model)trainer1.Train(avilaFNN, trainDataSet);
//trainDataSet = trainDataSet.SelectPermutation().ToArray();
////Console.WriteLine(evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
//listBox1.Items.Add(DateTime.Now + $" : 1st Training Completed, Accuracy is:{evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier())}");
//Trainer trainer3 = new Trainer(300, false, 0, new ExponentialDelayOptimizer(1.5, 0.9, 0.05));
//avilaFNN = (Model)trainer3.Train(avilaFNN, trainDataSet);
////Console.WriteLine(evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
//listBox1.Items.Add(DateTime.Now + $" : 2nd Training Completed, Accuracy is:{evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier())}");
//for (int i = 0; i <= 4; i++)
//{
// trainDataSet = trainDataSet.SelectPermutation().ToArray();
// //Trainer trainer2 = new Trainer(300, false, 0, new TriangularCyclicOptimizer(1000,1,0.05,0.95));
// Trainer trainer2 = new Trainer(200, false, 0, new TriangularCyclicOptimizer(1000,1,0.1,0.95));
// avilaFNN = (Model)trainer2.Train(avilaFNN, trainDataSet);
// //Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
// //Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
// listBox1.Items.Add(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier())}");
// var acc = evaluater.EvaluateInDetails(testDataSet2, avilaFNN);
// avilaFNN.SaveModel(accuracy: acc);
//}
for (int i = 0; i <= 14; i++)
{
trainDataSet = trainDataSet.SelectPermutation().ToArray();
//Trainer trainer2 = new Trainer(300, false, 0, new TriangularCyclicOptimizer(1000,1,0.05,0.95));
Trainer trainer2 = new Trainer(200, false, 0, new ExponentialDelayOptimizer(0.035, 0.9, 0.1));
avilaFNN = (Model)trainer2.Train(avilaFNN, trainDataSet);
//Console.WriteLine(avilaFNN.layers[2].Weight[0, 0]);
//Console.WriteLine($"第1轮第{i}次循环:" + evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier()));
listBox1.Items.Add(DateTime.Now + $" : {i + 1}th Training Completed , Accuracy is:{evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier())}");
var acc = evaluater.EvaluateInDetails(testDataSet2, avilaFNN);
avilaFNN.SaveModel(accuracy: acc);
}
trainDataSet = trainDataSet.SelectPermutation().ToArray();
Trainer trainer4 = new Trainer(100, false, 0, new ExponentialDelayOptimizer(0.02, 0.9, 0.1));
avilaFNN = (Model)trainer4.Train(avilaFNN, trainDataSet);
listBox1.Items.Add(DateTime.Now + $" : 22th Training Completed, Accuracy is:{evaluater.Evaluate(testDataSet1, avilaFNN, new AvilaClassifier())}");
var accuracy = evaluater.EvaluateInDetails(testDataSet2, avilaFNN);
avilaFNN.SaveModel(accuracy: accuracy);
}
