This project is no longer maintained!

Note: This is old readme file

Dependencies

PerformanceWork: https://github.com/faruknane/Performance-Work

IntelMKL 2020

DeepLearningFramework

Every expression is a term whose derivative can be calculated by the DeepLearningFramework automatically. Every Layer is consist of Terms which corresponds to sequential input. An example of XOR model is shown below.

Hyperparameters

Global, independent from model.

Hyperparameters.LearningRate = 0.2f;
Hyperparameters.Optimizer = new SGD();

Inputs/Arguments of Model

var x = new Input(2);
var y = new Input(1);

Creating a Model

var l1 = Layer.Dense(2, x, "sigmoid"); 
var model = Layer.Dense(1, l1, "sigmoid"); 

The loss/error function

var loss = Layer.SquaredError(model, y);

Training Process

Very simple If you manually assign inputs! Dot Minimize or Maximize whichever you want to use.

for (int epoch = 0; epoch < 1000; epoch++)
{
    x.SetSequenceLength(1);
    y.SetSequenceLength(1);
    x.SetInput(0, new float[2, 4] { { 1, 1, 0, 0 }, { 1, 0, 1, 0 } } );
    y.SetInput(0, new float[1, 4] { { 0, 1, 1, 0 } } );
    
    loss.Minimize();
    //Console.WriteLine("loss: " + loss.GetTerm(0).GetResult()[0]);
}

Print Results

loss.DeleteTerms();
var result = model.GetTerm(0).GetResult();
Console.WriteLine("Results: " + result[0] + ", " + result[1] + ", " + result[2] + ", " + result[3]);

Recurrent Layers

var l1 = new Recurrent(1, x, 
    (Layer h, Layer x) => 
    {
        return h + x;
    }
);

var l1 = new Recurrent(10, x,
    (Layer h, Layer x) =>
    {
        var WH = new Variable(h.D1, h.D1, x.SequenceLength, setzero: true);
        return WH * h + Layer.Dense(h.D1, x, "");
    }
);

To support me: My Patreon