static void Main(string[] args)
{
// TODO: Use ComputeSharp to run code on GPU
/*
* using ReadWriteBuffer<float> buffer = Gpu.Default.AllocateReadWriteBuffer<float>(1000);
* Gpu.Default.For(1000, id => buffer[id.X] = id.X);
* float[] array = buffer.GetData();
*/
var dataPath = FileHelper.FindAppFolder("Data");
int imageWidth, imageHeight;
byte[][] images;
byte[] labels;
{
var bytes = File.ReadAllBytes(Path.Combine(dataPath, "MNIST", "train-images.idx3-ubyte"));
var reader = new PacketReader();
reader.Open(bytes);
var magic = reader.ReadInt();
if (magic != 0x00000803)
{
throw new Exception("Wrong magic number");
}
var imageCount = reader.ReadInt();
Trace.WriteLine($"Found {imageCount} images.");
imageHeight = reader.ReadInt();
imageWidth = reader.ReadInt();
images = new byte[imageCount][];
for (var i = 0; i < imageCount; i++)
{
images[i] = reader.ReadBytes(imageHeight * imageWidth).ToArray();
}
}
{
var bytes = File.ReadAllBytes(Path.Combine(dataPath, "MNIST", "train-labels.idx1-ubyte"));
var reader = new PacketReader();
reader.Open(bytes);
var magic = reader.ReadInt();
if (magic != 0x00000801)
{
throw new Exception("Wrong magic number");
}
var itemCount = reader.ReadInt();
Trace.WriteLine($"Found {itemCount} labels.");
labels = new byte[itemCount];
for (var i = 0; i < itemCount; i++)
{
labels[i] = reader.ReadByte();
}
}
{
var displayedImageIndex = 0;
var layerSetup = new int[] { imageWidth *imageHeight, 600, 400, 200, 100, 10 };
var net = new NeuralNet(layerSetup);
var netInputs = new float[imageWidth * imageHeight];
var netLearningRate = 0.1f;
var netResult = -1;
void ResetNetwork()
{
net = new NeuralNet(layerSetup);
RunNetworkAndRenderImage(displayedImageIndex);
}
void RunNetwork(int imageIndex)
{
for (var i = 0; i < netInputs.Length; i++)
{
netInputs[i] = images[imageIndex][i] / (float)byte.MaxValue;
}
var output = net.Compute(netInputs);
Debug.Assert(output.Length == 10);
// Get the digit the network has seen
var max = 0f;
for (var i = 0; i < output.Length; i++)
{
if (output[i] > max)
{
max = output[i];
netResult = i;
}
}
var expectedOutput = new float[10];
expectedOutput[labels[imageIndex]] = 1f;
net.BackpropagateError(expectedOutput);
}
var windowWidth = 1280;
var windowHeight = 720;
var imageScale = 8;
var labelText = "";
var imageText = "";
var statusText = "";
SDL.SDL_Init(SDL.SDL_INIT_VIDEO);
SDL.SDL_CreateWindowAndRenderer(windowWidth, windowHeight, 0, out var window, out var renderer);
SDL.SDL_SetWindowTitle(window, "Neural");
var font = Font.LoadFromChevyRayFolder(renderer, Path.Combine(dataPath, "Fonts", "ChevyRay - Softsquare Mono"));
var fontStyle = new FontStyle(font)
{
Scale = 2, LetterSpacing = 1
};
var smallFontStyle = new FontStyle(font)
{
Scale = 1, LetterSpacing = 1
};
var imageTexture = IntPtr.Zero;
var imageSourceRect = new SDL.SDL_Rect {
x = 0, y = 0, w = imageWidth, h = imageHeight
};
var imageDestRect = new SDL.SDL_Rect {
x = windowWidth / 2 - (imageWidth / 2 * imageScale), y = windowHeight / 2 - (imageHeight / 2 * imageScale), w = imageWidth * imageScale, h = imageHeight * imageScale
};
void LoadImageForRender(int imageIndex)
{
if (imageTexture != IntPtr.Zero)
{
SDL.SDL_DestroyTexture(imageTexture);
}
var pixels = new byte[imageWidth * imageHeight * 4];
for (var i = 0; i < imageWidth * imageHeight; i++)
{
pixels[i * 4 + 0] = pixels[i * 4 + 1] = pixels[i * 4 + 2] = images[imageIndex][i];
pixels[i * 4 + 3] = 0xff;
}
unsafe
{
fixed(byte *pixelsPointer = pixels)
{
var surface = SDL.SDL_CreateRGBSurfaceWithFormatFrom((IntPtr)pixelsPointer, imageWidth, imageHeight, 24, imageWidth * 4, SDL.SDL_PIXELFORMAT_ARGB8888);
imageTexture = SDL.SDL_CreateTextureFromSurface(renderer, surface);
SDL.SDL_FreeSurface(surface);
}
}
labelText = $"Label: {labels[imageIndex]}, Network output: {netResult}";
imageText = $"({imageIndex}/{images.Length})";
}
void RunNetworkAndRenderImage(int imageIndex)
{
RunNetwork(imageIndex);
LoadImageForRender(imageIndex);
}
RunNetworkAndRenderImage(displayedImageIndex);
var running = true;
AutoMode autoMode = AutoMode.None;
var autoImageIndex = 0;
var correctlyPredicted = 0;
while (running)
{
while (running && SDL.SDL_PollEvent(out var @event) != 0)
{
switch (@event.type)
{
case SDL.SDL_EventType.SDL_QUIT:
running = false;
break;
case SDL.SDL_EventType.SDL_WINDOWEVENT:
switch (@event.window.windowEvent)
{
case SDL.SDL_WindowEventID.SDL_WINDOWEVENT_CLOSE:
running = false;
break;
}
break;
case SDL.SDL_EventType.SDL_KEYDOWN:
switch (@event.key.keysym.sym)
{
case SDL.SDL_Keycode.SDLK_RIGHT:
if (displayedImageIndex < images.Length - 1)
{
displayedImageIndex++; RunNetworkAndRenderImage(displayedImageIndex);
}
break;
case SDL.SDL_Keycode.SDLK_LEFT:
if (displayedImageIndex > 0)
{
displayedImageIndex--; RunNetworkAndRenderImage(displayedImageIndex);
}
break;
case SDL.SDL_Keycode.SDLK_p:
if (autoMode == AutoMode.Train)
{
autoMode = AutoMode.None;
}
else
{
autoMode = AutoMode.Train;
autoImageIndex = 0;
}
break;
/* case SDL.SDL_Keycode.SDLK_t:
* net.Train(netLearningRate, netInputs);
* RunNetworkAndRenderImage(displayedImageIndex);
* break; */
case SDL.SDL_Keycode.SDLK_c:
if (autoMode == AutoMode.Check)
{
autoMode = AutoMode.None;
}
else
{
autoMode = AutoMode.Check;
autoImageIndex = 0;
correctlyPredicted = 0;
}
break;
case SDL.SDL_Keycode.SDLK_r:
if (autoMode == AutoMode.None)
{
ResetNetwork();
statusText = "";
}
break;
}
break;
}
}
if (!running)
{
break;
}
SDL.SDL_SetRenderDrawColor(renderer, 0, 0, 0, 255);
SDL.SDL_RenderClear(renderer);
// Draw image
SDL.SDL_RenderCopy(renderer, imageTexture, ref imageSourceRect, ref imageDestRect);
// Draw network
{
var start = new Point(10, 10);
var size = 4;
var spacing = 2;
var columnOffset = -1;
var maxPerColumn = 50;
for (var i = 0; i < net.Layers.Length; i++)
{
var layer = net.Layers[i];
for (var j = 0; j < layer.Outputs.Length; j++)
{
if (j % maxPerColumn == 0)
{
columnOffset++;
}
var value = Math.Clamp(layer.Outputs[j], 0f, 1f);
SDL.SDL_SetRenderDrawColor(renderer, (byte)(0xff * (1f - value)), (byte)(0xff * value), 0x00, 0xff);
var rect = new SDL.SDL_Rect {
x = start.X + (i + columnOffset) * (size + spacing), y = start.Y + (j % maxPerColumn) * (size + spacing), w = size, h = size
};
SDL.SDL_RenderFillRect(renderer, ref rect);
}
}
}
// Draw info
fontStyle.DrawText(8, windowHeight - fontStyle.Size - 8, "Left/Right to navigate, P to train, C to check accuracy, R to reset network");
fontStyle.DrawText(windowWidth / 2 - fontStyle.MeasureText(labelText) / 2, imageDestRect.y + imageDestRect.h + 8, labelText);
fontStyle.DrawText(windowWidth - fontStyle.MeasureText(imageText) - 8, windowHeight - fontStyle.Size - 8, imageText);
fontStyle.DrawText(windowWidth - fontStyle.MeasureText(statusText) - 8, 8, statusText);
smallFontStyle.DrawText(8, windowHeight - fontStyle.Size * 4 - 12, "Written by Elisee Maurer (@elisee)");
smallFontStyle.DrawText(8, windowHeight - fontStyle.Size * 3 - 12, "MNIST database: yann.lecun.com/exdb/mnist/");
smallFontStyle.DrawText(8, windowHeight - fontStyle.Size * 2 - 12, "Font by Chevy Ray - pixel-fonts.com");
SDL.SDL_RenderPresent(renderer);
// Thread.Sleep(1);
switch (autoMode)
{
case AutoMode.Train:
for (var i = 0; i < 10; i++)
{
RunNetwork(autoImageIndex);
net.Train(netLearningRate, netInputs);
autoImageIndex++;
if (autoImageIndex == images.Length)
{
autoMode = AutoMode.None;
break;
}
}
statusText = $"Trained on {autoImageIndex} images.";
RunNetworkAndRenderImage(displayedImageIndex);
break;
case AutoMode.Check:
for (var i = 0; i < 10; i++)
{
RunNetwork(autoImageIndex);
if (labels[autoImageIndex] == netResult)
{
correctlyPredicted++;
}
autoImageIndex++;
if (autoImageIndex == images.Length)
{
autoMode = AutoMode.None;
break;
}
}
var percent = Math.Round(100f * correctlyPredicted / autoImageIndex, 2, MidpointRounding.AwayFromZero);
statusText = $"Correctly predicted {percent:0.00}% ({correctlyPredicted}/{autoImageIndex})";
break;
}
}
}
}