public JobCNNClassify(GRasterLayer featureRasterLayer, GRasterLayer labelRasterLayer, int epochs, int model, int width, int height, int channel)
{
_t = new Thread(() => {
ImageClassifyEnv env = new ImageClassifyEnv(featureRasterLayer, labelRasterLayer);
CNN cnn = new CNN(new int[] { channel, width, height }, env.ActionNum);
//training
Summary = "模型训练中";
for (int i = 0; i < epochs; i++)
{
int batchSize = cnn.BatchSize;
var(states, labels) = env.RandomEval(batchSize);
double[][] inputX = new double[batchSize][];
for (int j = 0; j < batchSize; j++)
{
inputX[j] = states[j];
}
double loss = cnn.Train(inputX, labels);
Process = (double)i / epochs;
}
//classify
Summary = "分类应用中";
IRasterLayerCursorTool pRasterLayerCursorTool = new GRasterLayerCursorTool();
pRasterLayerCursorTool.Visit(featureRasterLayer);
//GDI graph
Bitmap classificationBitmap = new Bitmap(featureRasterLayer.XSize, featureRasterLayer.YSize);
Graphics g = Graphics.FromImage(classificationBitmap);
//
int seed = 0;
int totalPixels = featureRasterLayer.XSize * featureRasterLayer.YSize;
//应用dqn对图像分类
for (int i = 0; i < featureRasterLayer.XSize; i++)
{
for (int j = 0; j < featureRasterLayer.YSize; j++)
{
//get normalized input raw value
double[] normal = pRasterLayerCursorTool.PickNormalValue(i, j);
//}{debug
double[] action = cnn.Predict(normal);
//convert action to raw byte value
int gray = env.RandomSeedKeys[NP.Argmax(action)];
//后台绘制，报告进度
Color c = Color.FromArgb(gray, gray, gray);
Pen p = new Pen(c);
SolidBrush brush = new SolidBrush(c);
g.FillRectangle(brush, new Rectangle(i, j, 1, 1));
//report progress
Process = (double)(seed++) / totalPixels;
}
}
//保存结果至tmp
string fullFileName = Directory.GetCurrentDirectory() + @"\tmp\" + DateTime.Now.ToFileTimeUtc() + ".png";
classificationBitmap.Save(fullFileName);
//complete
Summary = "CNN训练分类完成";
Complete = true;
OnTaskComplete?.Invoke(Name, fullFileName);
});
}
public void ClassificationByCNN()
{
//loss
double _loss = 1.0;
//training epochs
int epochs = 100;
//
GRasterLayer featureLayer = new GRasterLayer(featureFullFilename);
GRasterLayer labelLayer = new GRasterLayer(trainFullFilename);
//create environment for agent exploring
IEnv env = new ImageClassifyEnv(featureLayer, labelLayer);
//assume 18dim equals 3x6 (image)
CNN cnn = new CNN(new int[] { 1, 3, 6 }, env.ActionNum);
//training
for (int i = 0; i < epochs; i++)
{
int batchSize = cnn.BatchSize;
var(states, labels) = env.RandomEval(batchSize);
double[][] inputX = new double[batchSize][];
for (int j = 0; j < batchSize; j++)
{
inputX[j] = states[j];
}
_loss = cnn.Train(inputX, labels);
}
//in general, loss is less than 5
Assert.IsTrue(_loss < 5.0);
//apply cnn to classify featureLayer
IRasterLayerCursorTool pRasterLayerCursorTool = new GRasterLayerCursorTool();
pRasterLayerCursorTool.Visit(featureLayer);
//get normalized input raw value
double[] normal = pRasterLayerCursorTool.PickNormalValue(50, 50);
double[] action = cnn.Predict(normal);
int landCoverType = env.RandomSeedKeys[NP.Argmax(action)];
//do something as you need. i.e. draw landCoverType to bitmap at position ( i , j )
//the classification results are not stable because of the training epochs are too few.
Assert.IsTrue(landCoverType >= 0);
}
