go-qndnn

quick 'n' dirty neural network (for practical use)

Introduction

This package contains a simple Go implementation for neural networks; for practical everyday-use in common use-cases. It is neither heavily optimized to be the best package around, nor does this package contain an exhaustive variety of mathematical functions. It supports Sigmoid, Tanh and ReLU. It leverages Go primitives.

nn := qndnn.NewNeuralNet(nil, 4, 3, 3, 1) // sigmoid is default; input (4), hidden1 (3), hidden2 (3), output (1)
// qndnn.NewNeuralNet(qndnn.WithRelu(), 4, 3, 3, 1) // – to use with relu
// qndnn.NewNeuralNet(qndnn.WithTanh(), 4, 3, 3, 1) // - to use with tanh

// to retrieve output with input values
out, err := nn.Output([]float64{1, 2, 3, 4})

// to train on expectations
err = nn.Train(
	[]Expectation{
        {
            Input: []float64{1, 2, 3, 4},
            Output: []float64{.42},
        },
    }, 
	0.01, // learning rate
	qndnn.RoundStrategy(1000), // train for 1000 rounds; other options include ThresholdStrategy (see examples)
)

serializedBase64, err := nn.Serialize() // to serialize net (weights, biases)

nn, err = NewNeuralNetFromSerialized(nil, serializedBase64) // deserialize serialized net into usable structure; initialized with sigmoid 
//nn, err = NewNeuralNetFromSerialized(qndnn.WithRelu(), serializedBase64) // - to initialize with relu
//nn, err = NewNeuralNetFromSerialized(qndnn.WithTanh(), serializedBase64) // - to initialize with tanh