sabina

Sabina

Sabina is a machine learning library.
This gem provides tools for Multi-Layer Perceptrons and Auto-Encoders.

Installation

Add this line to your application's Gemfile:

gem 'sabina'

And then execute:

$ bundle

Or install it yourself as:

$ gem install sabina

Usage

MultilayerPerceptron

Example of normal usage is shown below.

require 'sabina'

DIM = 2
K = 3
EPOCH = 100

training_data = Sabina::MultilayerPerceptron.load_csv('training_data.csv')

options = {
  :layers => [
    Sabina::Layer::MPInputLayer.new(DIM),
    Sabina::Layer::MPHiddenLayer.new(8),
    Sabina::Layer::MPOutputLayer.new(K)
  ],
  :mini_batch_size => 10,
  :learning_rate => 0.01,
  :training_data => training_data,
}

mp = Sabina::MultilayerPerceptron.new(options)

EPOCH.times do |t|
  mp.learn
  error = mp.error(training_data)
  puts " error : #{error}"
end

AutoEncoder

Example of normal usage is shown below. Use SparseAutoEncoder class when the number of input units is less than that of hidden units.

require 'sabina'

DIM = 2
EPOCH = 100

original_data = Sabina::AutoEncoder.load_csv('training_data.csv')

options = {
  :layers => [
    Sabina::Layer::AEInputLayer.new(DIM),
    Sabina::Layer::AEHiddenLayer.new(8),
    Sabina::Layer::AEOutputLayer.new(DIM)
  ],
  :mini_batch_size => 10,
  :learning_rate => 0.01,
  :training_data => original_data,
}

sae = Sabina::SparseAutoEncoder.new(options)

EPOCH.times do |t|
  sae.learn
  error = sae.error(original_data)
  puts " error : #{error}"
end

About a training data CSV file format

Examples of a CSV file are shown below.

x0,x1,label
0.8616722150185228,0.7958526101017311,0
0.548524744634457,0.8355704092991548,1
0.2430915120750876,0.6252296416575435,1
0.968877668321639,0.7502385938940324,0
...

This is a example for two-dimensional vector data. For example, if you want to input D-dimensional vector data, write x0,x1,...,x(D-1),label at the first line. The column of label is used for a cluster id. For example, if there are three clusters in training data, a number at the label column will be 0, 1 or 2.

When you prepare a CSV file, load the file as shown below.

training_data = Sabina::MultilayerPerceptron.load_csv('training_data.csv')

When you use a auto-encoder, load a CSV file as shown below.

original_data = Sabina::AutoEncoder.load_csv('training_data.csv')

Configuration

You can set default values by using Sabina.configure method. These values could be overwritten by providing an argument.

Sabina.configure do |config|
  config.layers = [
    Sabina::Layer::MPInputLayer.new(2),
    Sabina::Layer::MPHiddenLayer.new(8),
    Sabina::Layer::MPOutputLayer.new(3)
  ]
  config.mini_batch_size = 10
  config.learning_rate = 0.01
  config.training_data = Sabina::MultilayerPerceptron.load_csv('training_data.csv')
end

options = {
  :mini_batch_size => 20
}

mp_01 = Sabina::MultilayerPerceptron.new
mp_02 = Sabina::MultilayerPerceptron.new(options)

mp_01.mini_batch_size # => 10
mp_02.mini_batch_size # => 20

Your own layer class

You can create your own layer class. In the following example below, a rectified linear function is set as an activation function. @f_ is differentiation of @f.

class MyHiddenLayer < Sabina::Layer::BaseLayer
  def initialize(size)
    super
    # f(x) = max(0, x)
    @f = ->(x){ x > 0.0 ? x : 0.0 }
    @f_ = ->(x){ x > 0.0 ? 1.0 : 0.0 }
  end
end

options = {
  :layers => [
    Sabina::Layer::MPInputLayer.new(DIM),
    MyHiddenLayer.new(16),
    MyHiddenLayer.new(8),
    Sabina::Layer::MPOutputLayer.new(K)
  ],
  :mini_batch_size => 10,
  :learning_rate => 0.01,
  :training_data => training_data,
}

Examples

These examples require gnuplot version 5.0 or later.

examples/example_mp_01/

Run examples/example_mp_01/main.rb.

examples/example_mp_02/

Run examples/example_mp_02/main.rb.

examples/example_ae_01/

Run examples/example_ae_01/main.rb.

MNIST Example

https://github.com/seinosuke/sabina_mnist_example

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/seinosuke/sabina. This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the Contributor Covenant code of conduct.

License

The gem is available as open source under the terms of the MIT License.