See https://github.com/arnab/ocr/blob/

weka_svm_classifier.rb

require "java"
require "./lib/weka.jar"

def read_data(filename)
  Java::WekaCoreConverters::ConverterUtils::DataSource.new(filename).get_data_set
end

def normalize(data)
  puts data.first.to_double_array.inspect
  normalizer = Java::WekaFiltersUnsupervisedAttribute::Normalize.new
  normalizer.set_input_format(data)
  data = Java::WekaFilters::Filter.use_filter(data, normalizer)
  puts data.first.to_double_array.inspect
  data
end

def train(algo_type, data)
  algo = Java::WekaClassifiersFunctions::LibSVM.new

  # for SVM
  options = [
             ['S', 0], # svm:type C-SVC
             ['K', 2], # kernel-type: linear
             ['D', 0], # degree
             ['G', 0], # gamma
             ['R', 0], # coef0
             ['C', 10], # Coeff C
             ['N', 0], # nu
             ['Z', 0], # do not normalize
             ['E', 0.001], # tolerance: eps
             ['H', 0], # do not use shrinking
             ['M', 1000] # cache memory, 1GB
            ]
  options = options.map {|o| "-#{o.join(' ')}"}
  algo.set_options(options)
  puts "building model with #{algo.class} on #{data.size} data points, with options: #{options.join(' ')}"

  algo.build_classifier(data)
  algo
end

def test(model, data)
  classified_attr_num = data.num_attributes - 1
  failures = []
  puts "Testing #{data.size} data points"
  data.each_with_index do |d, i|
    expected_class = d.to_string(classified_attr_num)
    res_class = model.classify_instance(d)
    d.set_class_value(res_class)
    found_class = d.to_string(classified_attr_num)
    failures << i unless expected_class == found_class
    if i % 1000 == 0
      puts "Accuracy at #{i}/#{data.size}: #{100 * (1 - failures.size.to_f / (i+1))} %"
    end
  end
  puts "Accuracy: #{100 * (1 - failures.size.to_f / data.size)} %"
end

training_data = read_data("data/img_train_data.csv")
test_data = read_data("data/img_test_data.csv")

# classify the last attr (the label)
[training_data, test_data].each do |dataset|
  dataset.set_class_index(dataset.num_attributes - 1)
end

training_data = normalize(training_data)
test_data = normalize(test_data)

model = train(:svm, training_data)
test(model, test_data)