puzzlebox.neuralnets's Introduction

PuzzleBox.NeuralNets

This library is in beta.

PuzzleBox.NeuralNets is a .NET library for building and training neural networks. It's designed mainly to be easy to use and extend. The features are prioritised according to the needs of another project, for analysing the audio of bee hives.

It was inspired by the fluent interface of ConvNetSharp, a port of ConvNetJS, but I wanted transposed convolutional layers which ConvNetSharp didn't support at the time. I found other libraries overly difficult to learn and use, and I wanted a deeper understanding of neural networks which is why I decided to write this library.

Examples

XOR

// Build
var net = new Net(2)  // Input size 2
    .Dense(2)         // Fully-connected layer with output size 2
    .Sigmoid()
    .Dense(1)         // Fully-connected layer with output size 1
    .Sigmoid();

// Train
var trainingData = new (Tensor input, Tensor output)[]
    {
        (new float[] { 0, 0 }, 0),
        (new float[] { 0, 1 }, 1),
        (new float[] { 1, 0 }, 1),
        (new float[] { 1, 1 }, 0),
    };

var trainer = new Trainer(net, 0.15f);
var cost = await trainer.TrainAsync(10000, trainingData);

// Run
var output = net.FeedForwards(input);

More Complex Example

This example is more involved than a truth table with values between 0 and 1 forming a circular shape.

CNN Autoencoder

var net = new Net(5, 5)                        // Input size 5x5
    .Convolution(new [] { 3, 3 }, 2)           // Two 3x3 kernels
    .Relu()
    .Dense(new Size(new [] { 1, 1 }))          // Compress input to single value
    .Relu()
    .ConvolutionTranspose(new[] { 5, 5 }, 2)   // Two 5x5 kernels
    .Sigmoid()
    .Dense(new Size(new[] { 5, 5 }))
    .Sigmoid();

Current Features

Layers
- Dense
- Convolution
- Transpose Convolution
Activation functions
- Sigmoid
- TanH
- Relu
Cost functions:
- Squared error
- Cross entropy

Possible Features

Layers
- CNN with bias
- N-dimensional CNN
- Transformer

Recommend Projects

jasonkstevens / puzzlebox.neuralnets Goto Github PK