neuralnet

neuralnet

Implements a basic L-Layer neural network. Current features are

• Inputs are images only (classification is based on folder location)
• Binary classification only
• Supports only relu, tanh and sigmoid activation functions
• Auto-initialize weights
• Augment data by flipping images horizontally
• Normalize data sets
• Use L2 regularization
• Use dropout
• Split training set into mini batches
• Use gradient descent with momentum

How to use

Define the hyper parameters

Define the L layer net by using the neuralnet.NewHyperParametersBuilder() with the following options

• AddLayers(a ActivationFuncName, neurons ...uint) - adds layers with the specified neurons and activation function
• AddNLayer(a ActivationFuncName, neurons uint, n uint) - adds n indentical layers to the net
• SetLearningRate(learningRate float64) - The learning rate to use, defaults to 0.01
• SetIterations(iterations uint) - number of iterations used to train the model, defaults to 1000
• SetRegularizationFactor(regularizationFactor float64) - the regularization factor to use. 0 indicates not to regularize.
• SetDropoutKeepProbability - enables dropout by specifying the probability neurons should be kept (i.e. not dropped). 0 indicates not to dropout.
• SetMiniBatchSize - splits the training set into mini batches for large data sets
• UseGradientDescentWithMomentum(beta float64) - uses a exponential moving average of gradients when minimizing, allowing the learning rate to be higher as it dampens out oscillations.

The last layer must be a single neuron using the sigmoid activation function for binary classification.

e.g.

hyperParams, err := neuralnet.NewHyperParametersBuilder().
    AddLayers(neuralnet.ActivationFuncNameReLU, 3, 2).
    AddLayer(neuralnet.ActivationFuncNameSigmoid, 1)
    SetLearningRate(0.15).
    SetIterations(5000).
    SetRegularizationFactor(0.5).
    SetDropoutKeepProbability(0.75).
    SetMiniBatchSize(1024).
    UseGradientDescentWithMomentum(0.9).
    Build()

Load the training set

Once defined, create a training set using the neuralnet.NewImageSetBuilder() with the following options

• WithPathPrefix(pathPrefix string) - defines a root folder to use
• AddFolder(pathToFolder string, classification bool) - adds a folder to the training set, with the classification to use
• AddImage(pathToImage string, classification bool) - adds a single image, with the classification to use
• ResizeImages(width, height uint) - resize all the images
• AugmentFlipHorizontal() - doubles the data set by considering the images flipped horizontally
• Normalize() - normalizes the data. Note that if the training set is normalized, the test set will also need to be normalized.

If the images are not being resized, they need to be all of the same height and width.

e.g.

trainingDataSet, err := neuralnet.NewImageSetBuilder().
    AugmentFlipHorizontal().
    WithPathPrefix("datasets/Vegetable Images/train").
    AddFolder("Cabbage", false).
    AddFolder("Carrot", true).
    ResizeImages(32, 32).
    Normalize().
    Build()

Train the model with dataset

model, err := hyperParams.TrainModel(trainingDataSet)

Verify accuracy of training set

model.Predict(trainingDataSet)

Load the testing set

e.g.

testDataSet, err := neuralnet.NewImageSetBuilder().
    WithPathPrefix("../datasets/Vegetable Images/test").
    AddFolder("Cabbage", false).
    AddFolder("Carrot", true).
    ResizeImages(32, 32).
    Normalize().
    Build()

Verify accuracy of test set

model.Predict(testDataSet)