// Return Imagesource of digit <datasetnumber> in MNIST data base
ImageSource GetImage(int datasetnumber)
{
// Generate an empty bitmap
SKBitmap bitmap = new SKBitmap(28, 28, SKColorType.Rgba8888, SKAlphaType.Premul);
// Generate pixel data from neural net input data which is already normalized to [0.01,1]
Vector <double> pixel = 255 - mnistdata.Input(datasetnumber).Subtract(0.01) / 0.99 * 255;
// Copy pixel data to bitmap
for (int y = 0; y < bitmap.Height; y++)
{
for (int x = 0; x < bitmap.Width; x++)
{
int rgb = (int)pixel[y * 28 + x];
bitmap.SetPixel(x, y, Color.FromRgb(rgb, rgb, rgb).ToSKColor());
}
}
// Comvert bitmap to data object (to access it as stream)
SKData data = SKImage.FromBitmap(bitmap).Encode();
// Read image from stream and return
return(ImageSource.FromStream(data.AsStream));
}
public TrainAndTestPage(NeuralNetRunType runtype, NeuralNet neuralnet, MNISTDataManager mnistdata)
{
// Define GUI
// Page properties
Title = string.Format("{0} Net", runtype.ToString().ToUpperFirstOnly());
BackgroundColor = Color.SteelBlue;
Padding = new Thickness(Application.Current.MainPage.Width * 0.05, Application.Current.MainPage.Height * 0.05);
// Label for description
Label description = new Label
{
Text = string.Format("Please define the number of data sets from the MNIST {0} data base to be used for {0}ing. Each data set represents one digit. The {0} data base holds about {1:N0} data sets in total.", runtype.ToString(), mnistdata.CountData),
Margin = new Thickness(0, 0, 0, Application.Current.MainPage.Height * 0.075)
};
// Label for number of training / testing sets used
Label label_useddatasets = new Label
{
Text = String.Format("{0:N0} data set{1} selected", mnistdata.UsedDataSets, mnistdata.UsedDataSets == 0 ? "" : "s"),
Margin = new Thickness(0, 0, 0, Application.Current.MainPage.Height * 0.075)
};
// Slider for number of data sets used for training / testing
Slider slider_useddatasets = new Slider(1, mnistdata.CountData, mnistdata.UsedDataSets);
slider_useddatasets.ValueChanged += (s, e) =>
{
mnistdata.UsedDataSets = (int)e.NewValue;
label_useddatasets.Text = String.Format("{0:N0} data set{1} selected", mnistdata.UsedDataSets, mnistdata.UsedDataSets == 0 ? "" : "s");
};
// Progress bar
ProgressBar progressbar = new ProgressBar {
Progress = 0
};
// Label for showing progress
Label label_progress = new Label
{
Text = "Currently no calculation running",
};
// Button to start training / testing
Button button = new Button
{
Text = string.Format("Start {0}ing", runtype.ToString().ToUpperFirstOnly()),
WidthRequest = Application.Current.MainPage.Width * 0.5,
HeightRequest = Application.Current.MainPage.Height * 0.10
};
// Define page
Content = new ScrollView
{
Content = new StackLayout
{
Children =
{
description,
label_useddatasets,
slider_useddatasets,
button,
progressbar,
label_progress
}
}
};
// Button event
// Here training and test sessions are executed
button.Clicked += async(s, e) =>
{
// Check if another calculation is currently running
if (activerun == true)
{
var answer = await DisplayAlert("Warning", "Do you want to stop current calculation?", "Yes", "No");
if (answer)
{
endrun = true;
}
return;
}
// Check if selected data set is big
if (mnistdata.UsedDataSets > mnistdata.UsedDataSetsWarning)
{
bool answer = await DisplayAlert("Warning", "Please be aware that big data sets effect run time. Continue?", "Yes", "No");
if (!answer)
{
return;
}
}
// Set activerun and endrun flags
activerun = true;
endrun = false;
// Disable Back button
NavigationPage.SetHasBackButton(this, false);
// Rename Button to Stop
button.Text = string.Format("Stop {0}ing", runtype.ToString().ToUpperFirstOnly());
// Define parameters for progress bar calculation (increase bar 100 times or each 20 runs)
int progresssteps = Math.Max(100, mnistdata.UsedDataSets / 20);
int progressspan = (int)Math.Ceiling((double)mnistdata.UsedDataSets / progresssteps);
// Counter for training / testing runs
int i = 0;
// Initialize progress bar
label_progress.Text = i.ToString() + " / " + mnistdata.UsedDataSets.ToString();
await progressbar.ProgressTo((double)i / mnistdata.UsedDataSets, 1, Easing.Linear);
// Train neural net
if (runtype == NeuralNetRunType.train)
{
while (i < mnistdata.UsedDataSets && endrun == false)
{
// Counter for runs in progress step
int j = 0;
// Training loop for one progress step done as task
await Task.Run(() =>
{
while (j < progressspan && i < mnistdata.UsedDataSets)
{
// Train net with data set i
neuralnet.Train(mnistdata.Input(i), mnistdata.Output(i));
j++;
i++;
}
// Remember number of training data sets
neuralnet.TrainingDataCounter += j;
});
// Update progress bar
label_progress.Text = i.ToString() + " / " + mnistdata.UsedDataSets.ToString();
await progressbar.ProgressTo((double)i / mnistdata.UsedDataSets, 1, Easing.Linear);
}
// Show mesage
await DisplayAlert("Result", string.Format("Neural net trained with {0:N0} data sets", i), "OK");
}
if (runtype == NeuralNetRunType.test)
{
// Setup scorecard for capturing test results
// digittotal counts all trained figures, digitcorrect the ones that are identified correctly
Vector <double> digittotal = Vector <double> .Build.Dense(10);
Vector <double> digitcorrect = Vector <double> .Build.Dense(10);
// Testing loop
while (i < mnistdata.UsedDataSets && endrun == false)
{
// Counter for runs in progress step
int j = 0;
// Testing loop for one progress step done as task
await Task.Run(() =>
{
while (j < progressspan && i < mnistdata.UsedDataSets)
{
// Increase counter for testet digit (0..9)
int number = mnistdata.Number(i);
digittotal[number]++;
// Ask net
Vector <double> answer = neuralnet.Query(mnistdata.Input(i));
// Check if it's correct
if (answer.AbsoluteMaximumIndex() == number)
{
digitcorrect[number]++;
}
j++;
i++;
}
// Remember number of training data sets
neuralnet.TrainingDataCounter += j;
});
// Update progress bar
label_progress.Text = i.ToString() + " / " + mnistdata.UsedDataSets.ToString();
await progressbar.ProgressTo((double)i / mnistdata.UsedDataSets, 1, Easing.Linear);
}
// Remeber performance result
neuralnet.Performance.Add(digitcorrect.Sum() / digittotal.Sum());
// Show test results
await Navigation.PushAsync(new ResultsMNISTPage(neuralnet, mnistdata, digittotal, digitcorrect));
}
// Reset progress bar
label_progress.Text = "Currently no calculation running";
await progressbar.ProgressTo(0.0, 250, Easing.Linear);
// Rename Button to Start
button.Text = string.Format("Start {0}ing", runtype.ToString().ToUpperFirstOnly());
// Enable Back button
NavigationPage.SetHasBackButton(this, true);
// Cancel activerun flag
activerun = false;
};
}