Micrograd.NET

A .NET idiomatic implementation of Andrej Karpathy's micrograd (see video here: video)

Example:

var model = new MLP(2, 16, 16, 1);
model.Parameters.Count().Should().Be(337);

var X = LoadCsv("x.csv", 2);
var y = LoadCsv("y.csv", 1);

(Value loss, double accuracy) Loss()
{
    var scores = X.Select(model.Eval).ToArray();

    // max-margin loss
    var losses = y.Zip(scores, (yi, si) => (1 - yi[0] * si[0]).ReLU()).ToArray();
    var dataLoss = losses.Aggregate((a, b) => a + b) * (1.0 / losses.Length);

    // L2 regularization
    var alpha = 1e-4;
    var regLoss = alpha * model.Parameters.Select(x => x * x).Aggregate((a, b) => a + b);
    var totalLoss = dataLoss + regLoss;

    // accuracy
    var accuracy = y.Zip(scores, (yi, si) => Math.Sign(yi[0].Data) == Math.Sign(si[0].Data))
        .ToArray();

    return (totalLoss, accuracy.Count(x => x) / (1.0 * accuracy.Length));
}

var (loss, acc) = Loss();
outputHelper.WriteLine($"{loss} {acc}");

// optimization
for (var k = 0; k < 100; k++)
{
    (loss, acc) = Loss();
    model.ZeroGradient();
    loss.Backward();

    var learningRate = 1.0 - 0.9 * k / 100.0;
    foreach (var p in model.Parameters)
        p.Update(learningRate);

    outputHelper.WriteLine($"step {k} loss {loss.Data}, accuracy {acc*100:0.00}%");
}

loss.Data.Should().BeApproximately(0.02, 0.02);
acc.Should().BeApproximately(1.0, 0.02);