tazdeb / ann

Machine Learning library for .NET Core

Ann

Machine Learning library for .NET Core.

Demo

Installation

PM> Install-Package Ann

Basic Usage

Configure a Network by defining the structure and meta-parametres

var network = new Network(LossFunctionType.CrossEntropy, new Flat(0.001), 5);

Creates a neural network object with a Cross-Entropy loss function, flat learning rate 0.001 and 5 output classes

Add layers to network configuration:

network.AddInputLayer(128, 1); 
network.AddConvolutionLayer(16, 5);
network.AddActivationLayer(ActivatorType.Relu); 
network.AddPoolingLayer(2); 
network.AddFlattenLayer();
network.AddDenseLayer(256, true);
network.AddSoftMaxLayer();

AddInputLayer(128, 1) input layer, which expect input with dimensions 128x128x1

AddConvolutionLayer(16, 5) convolution layer with 16 filters of size 5x5xD, where D is the depth of the output from a previous layer

AddActivationLayer(ActivatorType.Relu) activation layer with ReLU activation function

AddPoolingLayer(2) pooling layer with a vertical and horizontal stride 2

AddFlattenLayer() flattens the result

AddDenseLayer(256, true) fully connected layer with biases

AddSoftMaxLayer() SoftMax activation

Train Model

model.TrainModel(input, target);

First argument of the TrainModel() method accepts System.Array. The dimensions of the array must match with a input layer configuration. Second argument accepts bool[]. The length of the array must match number of classes provided to Network constructor.

Weights and biases will be adjasted using Stochastic Gradient Descent with backpropagation.

Save Model

After you are done with trainig you can save the model in JSON file for a later use:

var model = network.BuildModel();
model.Save("model.json");

Use Model

var model = new Model("model.json");
double[] prediction = model.Predict(input);

TrainModel() method accepts System.Array, the dimensions of the array must match with a input layer configuration.

Examples

There are two sample projects:

1. Ann.Mnist
2. Ann.Fingers

Authors

Pavel koryakin [email protected]

License

This project is licensed under the MIT License - see the LICENSE.md for details.

Acknowledgments

• Leon Bottou, Stochastic Gradient Descent Tricks
• Matt Mazur, A Step by Step Backpropagation Example
• Grant Sanderson, But what is a Neural Network?