public static void FindsABParamsUsingLevenbergMarquardtForDefaultSettings()
{
const float expectedA = 1.5769434603113077f;
const float expectedB = 0.8950608779109733f;
var(a, b) = Umap.FindABParams(1, 0.1f);
Assert.True(AreCloseEnough(a, expectedA));
Assert.True(AreCloseEnough(b, expectedB));
bool AreCloseEnough(float x, float y) => Math.Abs(x - y) < 0.01;
}
public static void FindsNearestNeighbors()
{
var nNeighbors = 10;
var umap = new Umap(random: new DeterministicRandomGenerator(42), numberOfNeighbors: nNeighbors);
var(knnIndices, knnDistances) = umap.NearestNeighbors(TestData, progress => { });
Assert.Equal(knnDistances.Length, TestData.Length);
Assert.Equal(knnIndices.Length, TestData.Length);
Assert.Equal(knnDistances[0].Length, nNeighbors);
Assert.Equal(knnIndices[0].Length, nNeighbors);
}
// Main
static void Main(string[] args)
{
// Download the MNIST data from Mr. YannAndréLeCun's web site
using (var wc = new WebClient())
{
if (!File.Exists(LABEL_FILE_NAME))
{
wc.DownloadFile(LABEL_FILE_URL, LABEL_FILE_NAME);
}
if (!File.Exists(IMAGE_FILE_NAME))
{
wc.DownloadFile(IMAGE_FILE_URL, IMAGE_FILE_NAME);
}
}
// Load the MNIST labels
var labels = GetLabels().ToArray();
// Load the MNIST image data
var pixels = GetImages().ToArray();
// Dimension reduction with UMAP
var umap = new Umap();
var epochs = umap.InitializeFit(pixels);
for (var i = 0; i < epochs; ++i)
{
umap.Step();
}
var embedding = umap.GetEmbedding().AsEnumerable();
// Convert the embedding into chart data
var graph = new Graph.Scatter()
{
x = embedding.Select((o) => o[0]),
y = embedding.Select((o) => o[1]),
text = labels.Select((o) => o.ToString()),
mode = "markers",
marker = new Graph.Marker {
color = labels, colorscale = "Rainbow", showscale = true
},
};
var chart = Chart.Plot(graph);
chart.WithTitle("MNIST Embedded via UMAP");
chart.WithXTitle("X");
chart.WithYTitle("Y");
chart.WithSize(800, 800);
chart.Show();
}
public static void StepMethod3D()
{
var umap = new Umap(random: new DeterministicRandomGenerator(42), dimensions: 3);
var nEpochs = umap.InitializeFit(TestData);
for (var i = 0; i < nEpochs; i++)
{
umap.Step();
}
var embedding = umap.GetEmbedding();
Assert.Equal(500, nEpochs);
AssertNestedFloatArraysEquivalent(TestResults3D, embedding);
}
static void Main()
{
// Note: The MNIST data here consist of normalized vectors (so the CosineForNormalizedVectors distance function can be safely used)
var data = MessagePackSerializer.Deserialize <LabelledVector[]>(File.ReadAllBytes("MNIST-LabelledVectorArray-60000x100.msgpack"));
data = data.Take(10_000).ToArray();
var timer = Stopwatch.StartNew();
var umap = new Umap(distance: Umap.DistanceFunctions.CosineForNormalizedVectors);
Console.WriteLine("Initialize fit..");
var nEpochs = umap.InitializeFit(data.Select(entry => entry.Vector).ToArray());
Console.WriteLine("- Done");
Console.WriteLine();
Console.WriteLine("Calculating..");
for (var i = 0; i < nEpochs; i++)
{
umap.Step();
if ((i % 10) == 0)
{
Console.WriteLine($"- Completed {i + 1} of {nEpochs}");
}
}
Console.WriteLine("- Done");
var embeddings = umap.GetEmbedding()
.Select(vector => new { X = vector[0], Y = vector[1] })
.ToArray();
timer.Stop();
Console.WriteLine("Time taken: " + timer.Elapsed);
// Fit the vectors to a 0-1 range (this isn't necessary if feeding these values down from a server to a browser to draw with Plotly because ronend because Plotly scales the axes to the data)
var minX = embeddings.Min(vector => vector.X);
var rangeX = embeddings.Max(vector => vector.X) - minX;
var minY = embeddings.Min(vector => vector.Y);
var rangeY = embeddings.Max(vector => vector.Y) - minY;
var scaledEmbeddings = embeddings
.Select(vector => new { X = (vector.X - minX) / rangeX, Y = (vector.Y - minY) / rangeY })
.ToArray();
const int width = 1600;
const int height = 1200;
using (var bitmap = new Bitmap(width, height))
{
using (var g = Graphics.FromImage(bitmap))
{
g.FillRectangle(Brushes.DarkBlue, 0, 0, width, height);
g.SmoothingMode = SmoothingMode.HighQuality;
g.TextRenderingHint = TextRenderingHint.ClearTypeGridFit;
g.InterpolationMode = InterpolationMode.HighQualityBicubic;
g.PixelOffsetMode = PixelOffsetMode.HighQuality;
using (var font = new Font("Tahoma", 6))
{
foreach (var(vector, uid) in scaledEmbeddings.Zip(data, (vector, entry) => (vector, entry.UID)))
{
g.DrawString(uid, font, Brushes.White, vector.X * width, vector.Y * height);
}
}
}
bitmap.Save("Output-Label.png");
}
var colors = "#006400,#00008b,#b03060,#ff4500,#ffd700,#7fff00,#00ffff,#ff00ff,#6495ed,#ffdab9"
.Split(',')
.Select(c => ColorTranslator.FromHtml(c))
.Select(c => new SolidBrush(c))
.ToArray();
using (var bitmap = new Bitmap(width, height))
{
using (var g = Graphics.FromImage(bitmap))
{
g.FillRectangle(Brushes.White, 0, 0, width, height);
g.SmoothingMode = SmoothingMode.HighQuality;
g.InterpolationMode = InterpolationMode.HighQualityBicubic;
g.PixelOffsetMode = PixelOffsetMode.HighQuality;
foreach (var(vector, uid) in scaledEmbeddings.Zip(data, (vector, entry) => (vector, entry.UID)))
{
g.FillEllipse(colors[int.Parse(uid)], vector.X * width, vector.Y * height, 5, 5);
}
}
bitmap.Save("Output-Color.png");
}
Console.WriteLine("Generated visualisation images");
Console.WriteLine("Press [Enter] to terminate..");
Console.ReadLine();
}