private BaseModel LoadModel(string modelPath, string weightsPath)
{
string text = File.ReadAllText(modelPath);
BaseModel model = Model.ModelFromJson(text);
model.LoadWeight(weightsPath);
model.Compile(loss: "categorical_crossentropy", optimizer: "adam", metrics: new string[] { "accuracy" });
return(model);
}
public void CreateModels()
{
var discriminatorOptimizer = new Adam(lr: 0.01f);
var ganOptimizer = new Adam(lr: 0.005f);
// Image generator
Sequential generatorSeq = new Sequential();
Input generatorInput = new Input(new Shape(ImageData.TotalInputShape));
generatorSeq.Add(generatorInput);
generatorSeq.Add(new Dense(128, activation: "relu"));
generatorSeq.Add(new Dropout(0.3f));
generatorSeq.Add(new BatchNormalization(momentum: 0.9f));
generatorSeq.Add(new Dense(256, activation: "relu"));
generatorSeq.Add(new BatchNormalization(momentum: 0.9f));
generatorSeq.Add(new Dense(512, activation: "relu"));
generatorSeq.Add(new BatchNormalization(momentum: 0.9f));
generatorSeq.Add(new Dense(2048, activation: "relu"));
generatorSeq.Add(new BatchNormalization(momentum: 0.9f));
generatorSeq.Add(new Dense(ImageData.OutputShape + ImageData.ManualInputShape, activation: "sigmoid"));
generatorModel = generatorSeq;
generatorModel.Compile(optimizer: ganOptimizer, loss: "mse", metrics: new string[] { "accuracy" });
// Image discriminator
Sequential discriminatorSeq = new Sequential();
discriminatorSeq.Add(new Dense(2048, activation: "relu", input_shape: new Shape(ImageData.OutputShape + ImageData.ManualInputShape)));
discriminatorSeq.Add(new Dropout(0.4f));
discriminatorSeq.Add(new Dense(512, activation: "relu"));
discriminatorSeq.Add(new Dropout(0.4f));
discriminatorSeq.Add(new Dense(128, activation: "relu"));
discriminatorSeq.Add(new Dropout(0.4f));
discriminatorSeq.Add(new Dense(32, activation: "relu"));
discriminatorSeq.Add(new Dropout(0.4f));
discriminatorSeq.Add(new Dense(1, activation: "sigmoid"));
discriminatorModel = discriminatorSeq;
discriminatorModel.Compile(optimizer: discriminatorOptimizer, loss: "binary_crossentropy", metrics: new string[] { "accuracy" });
// AdversarialModel
Sequential ganSeq = new Sequential();
//ganSeq.Add(new Concatenate(generatorModel.ToLayer(), generatorInput));
ganSeq.Add(generatorModel.ToLayer());
ganSeq.Add(discriminatorModel.ToLayer());
ganModel = ganSeq;
ganModel.Compile(optimizer: ganOptimizer, loss: "binary_crossentropy", metrics: new string[] { "accuracy" });
}
public static (BaseModel model, double accuracy) TrainAndTest(BaseModel model)
{
// Global parameters
string datasetPath = @"Sudoku.NeuralNetwork\Dataset\sudoku.csv.gz";
int numSudokus = 1000;
// ML parameters
double testPercent = 0.2;
float learningRate = .001F;
int batchSize = 32;
int epochs = 2;
Console.WriteLine("Initialize dataset");
var(sPuzzles, sSols) = DataSetHelper.ParseCSV(datasetPath, numSudokus);
var(_sPuzzzlesTrain, _sPuzzlesTest) = DataSetHelper.SplitDataSet(sPuzzles, testPercent);
var(_sSolsTrain, _sSolsTest) = DataSetHelper.SplitDataSet(sSols, testPercent);
Console.WriteLine("Preprocess data");
var sPuzzzlesTrain = DataSetHelper.PreprocessSudokus(_sPuzzzlesTrain);
var sSolsTrain = DataSetHelper.PreprocessSudokus(_sSolsTrain);
var sPuzzlesTest = DataSetHelper.PreprocessSudokus(_sPuzzlesTest);
var sSolsTest = DataSetHelper.PreprocessSudokus(_sSolsTest);
// Add optimizer
var adam = new Keras.Optimizers.Adam(learningRate);
model.Compile(loss: "sparse_categorical_crossentropy", optimizer: adam);
Console.WriteLine("Train model");
model.Fit(sPuzzzlesTrain, sSolsTrain, batch_size: batchSize, epochs: epochs);
// Test model
int correct = 0;
for (int i = 0; i < 1; i++)
{
Console.WriteLine("Testing " + i);
// Predict result
var prediction = Solve(sPuzzlesTest[i], model);
// Convert to sudoku
var sPredict = new Core.Sudoku()
{
Cells = prediction.flatten().astype(np.int32).GetData <int>().ToList()
};
var sSol = new Core.Sudoku()
{
Cells = ((sSolsTest[i] + 0.5) * 9).flatten().astype(np.int32).GetData <int>().ToList()
};
// Compare sudoku
var same = true;
for (int j = 0; j < 9; j++)
{
for (int k = 0; k < 9; k++)
{
if (sPredict.GetCell(j, k) != sSol.GetCell(j, k))
{
same = false;
}
}
}
Console.WriteLine("Prediction : \n" + sPredict);
Console.WriteLine("Solution : \n" + sSol);
if (same)
{
correct += 1;
}
}
// Calculate accuracy
var accuracy = (correct / sPuzzlesTest.size) * 100;
// Return
return(model, accuracy);
}
public async void TrainMovingData()
{
try {
// Compile
var optimizer = new Adam(lr: 0.001f);
model.Compile(optimizer: optimizer, loss: "mse", metrics: new string[] { "accuracy" });
// Train
MovingDataset[] movingTrainDatasets = LoadMovingDatasets();
int dataCount = 0;
List <float[, ]> inputList = new List <float[, ]>();
List <float[, ]> outputList = new List <float[, ]>();
for (int i = 0; i < movingTrainDatasets.Length; ++i)
{
MovingDataset datasetPart = movingTrainDatasets[i];
TrainData trainData = datasetPart.GetTrainData();
inputList.Add(trainData.inputs);
outputList.Add(trainData.outputs);
dataCount += trainData.inputs.GetLength(0);
}
float[,] inputs = new float[dataCount, MovingDataset.InputFloatCount];
float[,] outputs = new float[dataCount, MovingDataset.OutputFloatCount];
int inputIndex = 0;
foreach (float[,] input in inputList)
{
for (int i = 0; i < input.GetLength(0); ++i)
{
for (int f = 0; f < MovingDataset.InputFloatCount; ++f)
{
inputs[inputIndex, f] = input[i, f];
}
++inputIndex;
}
}
int outputIndex = 0;
foreach (float[,] output in outputList)
{
for (int i = 0; i < output.GetLength(0); ++i)
{
for (int f = 0; f < MovingDataset.OutputFloatCount; ++f)
{
outputs[outputIndex, f] = output[i, f];
}
++outputIndex;
}
}
for (; ;)
{
model.Fit(inputs, outputs, batch_size: 100, epochs: 100, verbose: 1);
SaveModel();
await Task.Delay(10);
}
//float[] predicts = model.Predict(x).GetData<float>();
//Stopwatch watch = new Stopwatch();
//watch.Start();
//for (int i = 0; i < 1; ++i) {
// predicts = model.Predict(x, verbose: 0).GetData<float>();
// //Debug.WriteLine($"Result: ({string.Join(",", predicts)})");
//}
//watch.Stop();
//MainWindow.Instance.Dispatcher.BeginInvoke(new Action(() => {
// MainWindow.Instance.DebugTextBox.Text = watch.GetElapsedMilliseconds().ToString();
//}));
} catch (Exception ex) {
MainWindow.Instance.Dispatcher.BeginInvoke(new Action(() => {
MainWindow.Instance.DebugTextBox.Text = ex.ToString();
}));
}
}
