void train_network()
{
var mnist = new Datasets.MNIST();
var trainSteps = mnist.train_images.Length / batch_size;
var validationSteps = mnist.test_images.Length / batch_size;
var wrapped_trainGenerator = new InMemoryMiniBatchGenerator(
new float[][][] { mnist.train_images, mnist.train_labels },
new C.Variable[] { imageVariable, categoricalVariable },
batch_size, true, false, "training");
var trainGenerator = new MultiThreadedGenerator(4, wrapped_trainGenerator);
var wrapped_testGenerator = new InMemoryMiniBatchGenerator(
new float[][][] { mnist.test_images, mnist.test_labels },
new C.Variable[] { imageVariable, categoricalVariable },
batch_size, true, false, "testing");
var validationGenerator = new MultiThreadedGenerator(4, wrapped_testGenerator);
var epochs = 6;
Model.fit_generator(
network,
learner,
trainer,
evaluator,
batch_size,
epochs,
trainGenerator, trainSteps,
validationGenerator, validationSteps,
computeDevice, "mnist_");
trainGenerator.Dispose(); trainGenerator = null;
validationGenerator.Dispose(); validationGenerator = null;
}
bool show_translations(bool use_network)
{
if (use_network)
{
network = load_best_network();
}
var mnist = new Datasets.MNIST();
var image_size = new int[] { input_shape[0], input_shape[1] };
var allData = new Data(mnist.train_images, image_size).All();
for (int index = 0; index < 1000; index++)
{
var img = allData[0][index];
var T = allData[1][index];
var img_t = allData[2][index];
var label = mnist.train_labels[index];
var digit = Util.argmax(label);
var mat = convertFloatRowTo2D(img, 28, 28);
var window_name = $"mnist_{index}_Digit_{digit}";
// allow the window to be resized
OpenCvSharp.Cv2.NamedWindow(window_name, OpenCvSharp.WindowMode.Normal);
OpenCvSharp.Cv2.ImShow(window_name, mat);
var mat_t = convertFloatRowTo2D(img_t, 28, 28);
var window_name_T = $"{T[0]}_{T[1]}";
OpenCvSharp.Cv2.NamedWindow(window_name_T, OpenCvSharp.WindowMode.Normal);
OpenCvSharp.Cv2.ImShow(window_name_T, mat_t);
var predicted_mat_t = new OpenCvSharp.Mat();
if (network != null)
{
var prediction = Model.predict(network, new float[][] { img, T }, computeDevice);
var predicted_img_t = prediction[0].ToArray();
predicted_mat_t = convertFloatRowTo2D(predicted_img_t, 28, 28);
var window_name_predicted_T = $"predicted_{T[0]}_{T[1]}";
OpenCvSharp.Cv2.NamedWindow(window_name_predicted_T, OpenCvSharp.WindowMode.Normal);
OpenCvSharp.Cv2.ImShow(window_name_predicted_T, predicted_mat_t);
}
OpenCvSharp.Cv2.WaitKey();
OpenCvSharp.Cv2.DestroyAllWindows();
//OpenCvSharp.Cv2.DestroyWindow(window_name);
//OpenCvSharp.Cv2.DestroyWindow(window_name_T);
GC.KeepAlive(mat);
GC.KeepAlive(mat_t);
GC.KeepAlive(predicted_mat_t);
}
return(false);
}
void run()
{
//if ( show_translations(use_network: true)==false ) { return; }
Console.Title = "Transforming AutoEncoders";
Console.WriteLine("Using: " + computeDevice.AsString());
create_network();
var mnist = new Datasets.MNIST();
var image_size = new int[] { input_shape[0], input_shape[1] };
var train_data = new Data(mnist.train_images, image_size);
var test_data = new Data(mnist.test_images, image_size);
train_network(train_data, test_data);
}
void train_network()
{
var mnist = new Datasets.MNIST();
var trainSteps = mnist.train_images.Length / batch_size;
var validationSteps = mnist.test_images.Length / batch_size;
var trainGenerator = new InMemoryMiniBatchGenerator(
new float[][][] { mnist.train_images, mnist.train_labels },
new C.Variable[] { imageVariable, categoricalLabel },
batch_size, shuffle: true, only_once: false, name: "train");
var mtTrainGenerator = new MultiThreadedGenerator(workers: 4, generator: trainGenerator);
var validationGenerator = new InMemoryMiniBatchGenerator(
new float[][][] { mnist.test_images, mnist.test_labels },
new C.Variable[] { imageVariable, categoricalLabel },
batch_size, shuffle: false, only_once: false, name: "validation");
var mtValidationGenerator = new MultiThreadedGenerator(workers: 4, generator: validationGenerator);
var epochs = 6;
Model.fit_generator(
network,
learner,
trainer,
evaluator,
batch_size,
epochs,
mtTrainGenerator, trainSteps,
mtValidationGenerator, validationSteps,
computeDevice,
prefix: "capsules_",
trainingLossMetricName: "Training Loss",
trainingEvaluationMetricName: "Training Error",
validationMetricName: "Validation Error");
mtTrainGenerator.Dispose();
mtValidationGenerator.Dispose();
}
