private void Form4_2_Load(object sender, EventArgs e)
{
cnn = new Cnn();
#region LeNet-5 结构
/*
* cnn.AddCnnConvolutionLayer(6, 32, 32, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
* 2, 2, CnnPooling.PoolingTypes.MaxPooling, false);
* cnn.AddCnnConvolutionLayer(16, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
* 2, 2, CnnPooling.PoolingTypes.MeanPooling, false, false);
* cnn.AddCnnConvolutionLayer(120, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
* 0, 0, CnnPooling.PoolingTypes.None, false, false);
* cnn.AddCnnFullLayer(84, CnnNode.ActivationFunctionTypes.Tanh, false);
* cnn.AddCnnFullLayer(10, CnnNode.ActivationFunctionTypes.Tanh, false);
* //*/
#endregion
cnn.AddCnnConvolutionLayer(6, 254 * 2, 252, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MaxPooling, false);
cnn.AddCnnConvolutionLayer(16, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MeanPooling, false, false);
cnn.AddCnnConvolutionLayer(20, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MeanPooling, false, false);
cnn.AddCnnFullLayer(84, CnnNode.ActivationFunctionTypes.Tanh, false);
cnn.AddCnnFullLayer(1, CnnNode.ActivationFunctionTypes.Tanh, false);
}
private void button2_Click(object sender, EventArgs e)
{
//if (args.length < 5)
//{
// System.out
// .println("Usage: \n\t-train trainfile\n\t-test predictfile\n\t-sep separator, default:','\n\t-eta eta, default:0.5\n\t-iter iternum, default:5000\n\t-out outputfile");
// return;
//}
String trainfile = @".\data\train.txt";
String testfile = @".\data\test.txt";
String outputfile = "outputfile.txt";
float eta = 0.05f;
int nIter = 1000;
List <CnnDemo.BP.DataNode> trainList = GetDataList(trainfile);
List <CnnDemo.BP.DataNode> testList = GetDataList(testfile);
StreamWriter sw = new StreamWriter(outputfile);
int typeCount = 3;
Cnn tmpCnn = new Cnn();
tmpCnn.AddCnnFullLayer(4, 14, CnnNode.ActivationFunctionTypes.Tanh, false);
tmpCnn.AddCnnFullLayer(20, CnnNode.ActivationFunctionTypes.Tanh, false);
tmpCnn.AddCnnFullLayer(10, CnnNode.ActivationFunctionTypes.Tanh, false);
tmpCnn.AddCnnFullLayer(10, CnnNode.ActivationFunctionTypes.Tanh, false);
tmpCnn.AddCnnFullLayer(3, CnnNode.ActivationFunctionTypes.Tanh, false);
for (int i = 0; i < nIter; i++)
{
foreach (var t in trainList)
{
tmpCnn.TrainFullLayer(t.getAttribList2().ToArray(), t.getTypes(), eta);
}
}
for (int i = 0; i < testList.Count(); i++)
{
CnnDemo.BP.DataNode test = testList[i];
double[] type = tmpCnn.PredictFullLayer(test.getAttribList2().ToArray());
List <float> attribs = test.getAttribList();
for (int n = 0; n < attribs.Count(); n++)
{
sw.Write(attribs[n] + ",");
}
double maxtype = type[0], max = 0;
for (int n = 0; n < 3; n++)
{
if (maxtype < type[n])
{
max = n;
maxtype = type[n];
}
}
sw.WriteLine(GetTypeName(max));
}
sw.Close();
}
private void CreateBP()
{
cnn = new Cnn();
cnn.AddCnnConvolutionLayer(8, bpWidth, bpHeight, 20, 20, 5, 5, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MaxPooling, false);
cnn.AddCnnConvolutionLayer(20, 10, 10, 3, 3, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MeanPooling, false, false);
cnn.AddCnnConvolutionLayer(40, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MeanPooling, false, false);
cnn.AddCnnConvolutionLayer(60, 5, 5, 1, 1, CnnNode.ActivationFunctionTypes.Tanh,
2, 2, CnnPooling.PoolingTypes.MeanPooling, false, false);
//cnn.AddCnnConvolutionLayer(80, 5, 5, 1, 1, 1, 2, 2, 1);
//cnn.AddCnnConvolutionLayer(100, 5, 5, 1, 1, 1, 2, 2, 1);
cnn.AddCnnFullLayer(300, CnnNode.ActivationFunctionTypes.Tanh, false);
cnn.AddCnnFullLayer(bpRectangleCount * 4, CnnNode.ActivationFunctionTypes.Tanh, false);
}
