ruby-em_algorithm 0.0.2

data/lib/em_algorithm/models/gaussian.rb

@@ -0,0 +1,47 @@
+module EMAlgorithm
+  class Gaussian < Model
+    attr_accessor :mu, :sigma, :dim
+
+    def initialize(mu = 0.0, sigma = 1.0)
+      @mu = mu
+      @sigma = sigma
+    end
+
+    def probability_density_function(x)
+      exp(-((x-@mu)**2.0)/(2.0*@sigma**2))/(sqrt(2.0*PI)*@sigma)
+    end
+
+    def probability_density_function_with_observation_weight(x_with_weight)
+      x = x_with_weight[0]
+      observation_weight = x_with_weight[1]
+      observation_weight * probability_density_function(x)
+    end
+
+    def update_average!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(0.0) do |sum, di|
+        sum + temp_weight_per_datum[di] * data_array[di]
+      end
+      @mu = data_sum / temp_weight
+    end
+
+    def update_sigma!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(0.0) do |sum, di|
+        sum + temp_weight_per_datum[di] * (data_array[di] - @mu) ** 2
+      end
+      @sigma = sqrt(data_sum / temp_weight)
+    end
+
+    def update_parameters!(data_array, temp_weight, temp_weight_per_datum)
+      update_average!(data_array, temp_weight, temp_weight_per_datum)
+      update_sigma!(data_array, temp_weight, temp_weight_per_datum)
+    end
+
+    def to_gnuplot
+      "exp(-((x-(#{@mu.round(DIGIT)}))**2.0)/(2.0*#{(@sigma ** 2).round(DIGIT)}))/(sqrt(2.0*pi)*#{@sigma.round(DIGIT)})"
+    end
+
+    def to_gnuplot_with_title(weight)
+      to_gnuplot + " w l axis x1y2 lw 3 title '#{weight.round(DIGIT)}*N(#{@mu.round(DIGIT)},#{@sigma.round(DIGIT)})'"
+    end
+  end
+end

data/lib/em_algorithm/models/gaussian.rb~

@@ -0,0 +1,47 @@
+module EMAlgorithm
+  class Gaussian < Model
+    attr_accessor :mu, :sigma, :dim
+
+    def initialize(mu = 0.0, sigma = 1.0)
+      @mu = mu
+      @sigma = sigma
+    end
+
+    def probability_density_function(x)
+      exp(-((x-@mu)**2.0)/(2.0*@sigma**2))/(sqrt(2.0*PI)*@sigma)
+    end
+
+    def probability_density_function_with_observation_weight(x_with_weight)
+      x = x_with_weight[0]
+      observation_weight = x_with_weight[1]
+      observation_weight * probability_density_function(x)
+    end
+
+    def update_average!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(0.0) do |sum, di|
+        sum + temp_weight_per_datum[di] * data_array[di]
+      end
+      @mu = data_sum / temp_weight
+    end
+
+    def update_sigma!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(0.0) do |sum, di|
+        sum + temp_weight_per_datum[di] * (data_array[di] - @mu) ** 2
+      end
+      @sigma = sqrt(data_sum / temp_weight)
+    end
+
+    def update_parameters!(data_array, temp_weight, temp_weight_per_datum)
+      update_average!(data_array, temp_weight, temp_weight_per_datum)
+      update_sigma!(data_array, temp_weight, temp_weight_per_datum)
+    end
+
+    def to_gnuplot
+      "exp(-((x-(#{@mu.round(DIGIT)}))**2.0)/(2.0*#{(@sigma ** 2).round(DIGIT)}))/(sqrt(2.0*pi)*#{@sigma.round(DIGIT)})"
+    end
+
+    def to_gnuplot_with_title(weight)
+      to_gnuplot + " w l axis x1y2 lw 3 title '#{weight.round(DIGIT)}*N(#{@mu.round(DIGIT)},#{@sigma.round(DIGIT)})'"
+    end
+  end
+end

data/lib/em_algorithm/models/md_gaussian.rb

@@ -0,0 +1,67 @@
+module EMAlgorithm
+  class MdGaussian < Model
+    attr_accessor :mu, :sigma2
+
+    def initialize(mu = GSL::Vector[0.0, 0.0], sigma2 = GSL::Matrix[[1.0, 0.0], [0.0, 1.0]])
+      # check mu
+      if mu.class != GSL::Vector
+        raise ArgumentError, "mu should be GSL::Vector."
+      end
+      @mu = mu
+      # check sigma2
+      if sigma2.class != GSL::Matrix
+        raise ArgumentError, "sigma2 should be GSL::Matrix."
+      elsif sigma2.size1 != @mu.size || sigma2.size2 != @mu.size
+        raise ArgumentError, "The size of sigma2 matrix does not match with mu vector."
+      end
+      @sigma2 = sigma2
+      @sqrt_sigma2_det = sqrt(@sigma2.det)
+      @sigma2_invert = @sigma2.invert
+    end
+
+    def probability_density_function(x)
+      exp(-((x-@mu) * @sigma2_invert * (x-@mu).trans)/2.0)/((sqrt(2.0*PI)**@mu.size)*@sqrt_sigma2_det)
+    end
+
+    def update_average!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(GSL::Vector.alloc(@mu.size).set_zero) do |sum, di|
+        sum + temp_weight_per_datum[di] * data_array[di]
+      end
+      @mu = data_sum / temp_weight
+    end
+
+    def update_sigma2!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(0.0) do |sum, di|
+        sum + temp_weight_per_datum[di] * (data_array[di] - @mu).trans * (data_array[di] - @mu)
+      end
+      @sigma2 = (data_sum / temp_weight)
+      @sqrt_sigma2_det = sqrt(@sigma2.det)
+      @sigma2_invert = @sigma2.invert
+    end
+
+    def update_parameters!(data_array, temp_weight, temp_weight_per_datum)
+      update_average!(data_array, temp_weight, temp_weight_per_datum)
+      update_sigma2!(data_array, temp_weight, temp_weight_per_datum)
+    end
+
+    def to_gnuplot
+      if @mu.size == 2
+        # [x - mu_x, y - mu_y] * [[s_x, s_xy], [s_xy,  s_y]] * [x - mu_x, y - mu_y]
+        # = s_x*(x - mu_x)**2 + 2*s_xy*(x - mu_x)(y - mu_y) + s_y*(y - mu_y)**2
+        sigma2_xy = @sigma2[0,1] + @sigma2[1,0]
+        xy = "+(#{sigma2_xy.round((DIGIT))})*(x-(#{@mu[0].round((DIGIT))}))*(y-(#{@mu[1].round((DIGIT))}))" if sigma2_xy > 0 || sigma2_xy < 0
+        "exp(-((#{@sigma2[0,0].round((DIGIT))})*(x-(#{@mu[0].round((DIGIT))}))**2.0+(#{@sigma2[1,1].round((DIGIT))})*(y-(#{@mu[1].round((DIGIT))}))**2.0#{xy})/2.0)/((sqrt(2.0*pi))**#{@mu.size}*(#{@sqrt_sigma2_det.round((DIGIT))}))"
+      else
+        "N(#{@mu.to_a.inspect}, #{@sigma2.to_a.inspect})"
+      end
+    end
+
+    def to_gnuplot_with_title(weight)
+      if @mu.size == 2
+        to_gnuplot + " w l lw 3 title '#{weight.round((DIGIT))}*N(#{@mu.map{|mu| mu.round((DIGIT))}.to_a.inspect},#{@sigma2.map{|sigma2| sigma2.round((DIGIT))}.to_a.inspect})'"
+      else
+        "N(#{@mu.to_a.inspect}, #{@sigma2.to_a.inspect})"
+      end
+    end
+  end
+end

data/lib/em_algorithm/models/md_gaussian.rb~

@@ -0,0 +1,67 @@
+module EMAlgorithm
+  class MdGaussian < Model
+    attr_accessor :mu, :sigma2
+
+    def initialize(mu = GSL::Vector[0.0, 0.0], sigma2 = GSL::Matrix[[1.0, 0.0], [0.0, 1.0]])
+      # check mu
+      if mu.class != GSL::Vector
+        raise ArgumentError, "mu should be GSL::Vector."
+      end
+      @mu = mu
+      # check sigma2
+      if sigma2.class != GSL::Matrix
+        raise ArgumentError, "sigma2 should be GSL::Matrix."
+      elsif sigma2.size1 != @mu.size || sigma2.size2 != @mu.size
+        raise ArgumentError, "The size of sigma2 matrix does not match with mu vector."
+      end
+      @sigma2 = sigma2
+      @sqrt_sigma2_det = sqrt(@sigma2.det)
+      @sigma2_invert = @sigma2.invert
+    end
+
+    def probability_density_function(x)
+      exp(-((x-@mu) * @sigma2_invert * (x-@mu).trans)/2.0)/((sqrt(2.0*PI)**@mu.size)*@sqrt_sigma2_det)
+    end
+
+    def update_average!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(GSL::Vector.alloc(@mu.size).set_zero) do |sum, di|
+        sum + temp_weight_per_datum[di] * data_array[di]
+      end
+      @mu = data_sum / temp_weight
+    end
+
+    def update_sigma2!(data_array, temp_weight, temp_weight_per_datum)
+      data_sum = (0..(data_array.size-1)).inject(0.0) do |sum, di|
+        sum + temp_weight_per_datum[di] * (data_array[di] - @mu).trans * (data_array[di] - @mu)
+      end
+      @sigma2 = (data_sum / temp_weight)
+      @sqrt_sigma2_det = sqrt(@sigma2.det)
+      @sigma2_invert = @sigma2.invert
+    end
+
+    def update_parameters!(data_array, temp_weight, temp_weight_per_datum)
+      update_average!(data_array, temp_weight, temp_weight_per_datum)
+      update_sigma2!(data_array, temp_weight, temp_weight_per_datum)
+    end
+
+    def to_gnuplot
+      if @mu.size == 2
+        # [x - mu_x, y - mu_y] * [[s_x, s_xy], [s_xy,  s_y]] * [x - mu_x, y - mu_y]
+        # = s_x*(x - mu_x)**2 + 2*s_xy*(x - mu_x)(y - mu_y) + s_y*(y - mu_y)**2
+        sigma2_xy = @sigma2[0,1] + @sigma2[1,0]
+        xy = "+(#{sigma2_xy.round((DIGIT))})*(x-(#{@mu[0].round((DIGIT))}))*(y-(#{@mu[1].round((DIGIT))}))" if sigma2_xy > 0 || sigma2_xy < 0
+        "exp(-((#{@sigma2[0,0].round((DIGIT))})*(x-(#{@mu[0].round((DIGIT))}))**2.0+(#{@sigma2[1,1].round((DIGIT))})*(y-(#{@mu[1].round((DIGIT))}))**2.0#{xy})/2.0)/((sqrt(2.0*pi))**#{@mu.size}*(#{@sqrt_sigma2_det.round((DIGIT))}))"
+      else
+        "N(#{@mu.to_a.inspect}, #{@sigma2.to_a.inspect})"
+      end
+    end
+
+    def to_gnuplot_with_title(weight)
+      if @mu.size == 2
+        to_gnuplot + " w l lw 3 title '#{weight.round((DIGIT))}*N(#{@mu.map{|mu| mu.round((DIGIT))}.to_a.inspect},#{@sigma2.map{|sigma2| sigma2.round((DIGIT))}.to_a.inspect})'"
+      else
+        "N(#{@mu.to_a.inspect}, #{@sigma2.to_a.inspect})"
+      end
+    end
+  end
+end

data/lib/em_algorithm/models/mixture.rb

@@ -0,0 +1,122 @@
+module EMAlgorithm
+  class Mixture < Model
+    attr_accessor :models, :weights
+
+    def initialize(options)
+      opts = {
+        :models =>
+        [
+          Gaussian.new(0.0, 9.0), Gaussian.new(10.0, 9.0)
+        ],
+        :weights =>
+        [
+          0.5, 0.5
+        ]
+      }.merge(options)
+      @models = opts[:models]
+      @weights = opts[:weights]
+      if !proper_weights?
+        argument_error
+      end
+      @temp_weights = Array.new(@models.size)
+      @temp_weight_per_datum = Array.new(@models.size).map { Array.new }
+    end
+
+    def argument_error
+      raise ArgumentError, "The summation of @weights must be equal to 1.0."
+    end
+
+    def add(model = Gaussian.new(0.0, 9.0), weight = 0.0)
+      @models << model
+      @weights << weight
+      if !proper_weights?
+        argument_error
+      end
+    end
+
+    def proper_weights?
+      @weights.inject(0) {|sum, v| sum + v} == 1.0
+    end
+
+    def probability_density_function(x)
+      pdf = 0.0
+      @models.each_with_index do |model, mi|
+        pdf += model.pdf(x) * @weights[mi]
+      end
+      pdf
+    end
+
+    def clear_temp_weight_per_datum!
+      @temp_weight_per_datum.each {|w| w.clear}
+    end
+
+    def calculate_posterior_data_array(data_array)
+      posterior_data_array = Array.new(data_array.size, 0.0)
+      @models.each_with_index do |model, mi|
+        data_array.each_with_index do |x, di|
+          posterior_data_array[di] += @weights[mi] * model.pdf(x)
+        end
+      end
+      posterior_data_array
+    end
+
+    def update_temp_weights!(data_array, posterior_data_array)
+      @models.each_with_index do |model, mi|
+        data_array.each_with_index do |x, di|
+          temp_weight_per_datum = @weights[mi] * model.pdf(x) / posterior_data_array[di]
+          temp_weight_per_datum = 0.0 if temp_weight_per_datum.nan?
+          @temp_weight_per_datum[mi] <<  temp_weight_per_datum
+        end
+        @temp_weights[mi] = @temp_weight_per_datum[mi].inject(0.0) {|sum, w| sum + w}
+      end
+    end
+
+    def update_weights!(data_array)
+      (0..(@models.size-1)).each do |mi|
+        @weights[mi] = @temp_weights[mi] / data_array.size
+      end
+    end
+
+    def update_parameters!(data_array)
+      @models.each_with_index do |model, mi|
+        model.update_parameters!(data_array, @temp_weights[mi], @temp_weight_per_datum[mi])
+      end
+      update_weights!(data_array)
+    end
+
+    # output types
+    # :full (default)
+    # :separate_only
+    # :mixture_only
+    def to_gnuplot(type = :full)
+      # output each model (currently assume Gaussian)
+      output = []
+      @models.each_with_index do |model, mi|
+        output << "#{@weights[mi].round(DIGIT)} * #{model.to_gnuplot_with_title(@weights[mi])}"
+      end
+      separate = output.join(", ")
+      # output mixture model (currently assume Gaussian Mixture model)
+      output = []
+      @models.each_with_index do |model, mi|
+        output << "#{@weights[mi].round(DIGIT)} * #{model.to_gnuplot}"
+      end
+      mixture = output.join(" + ")
+      case type
+      when :separate_only
+        return separate
+      when :mixture_only
+        return mixture
+      end
+      "#{separate}, #{mixture}"
+    end
+
+    def debug_output
+      <<-DEBUG_OUT
+  @weights=#{@weights.inspect}
+  @temp_weights=#{@temp_weights.inspect}
+  @models
+   #{@models.inspect}
+      DEBUG_OUT
+    end
+  end
+end

data/lib/em_algorithm/models/mixture.rb~

@@ -0,0 +1,122 @@
+module EMAlgorithm
+  class Mixture < Model
+    attr_accessor :models, :weights, :const
+
+    def initialize(options)
+      opts = {
+        :models =>
+        [
+          Gaussian.new(0.0, 9.0), Gaussian.new(10.0, 9.0)
+        ],
+        :weights =>
+        [
+          0.5, 0.5
+        ]
+      }.merge(options)
+      @models = opts[:models]
+      @weights = opts[:weights]
+      if !proper_weights?
+        argument_error
+      end
+      @temp_weights = Array.new(@models.size)
+      @temp_weight_per_datum = Array.new(@models.size).map { Array.new }
+    end
+
+    def argument_error
+      raise ArgumentError, "The summation of @weights must be equal to 1.0."
+    end
+
+    def add(model = Gaussian.new(0.0, 9.0), weight = 0.0)
+      @models << model
+      @weights << weight
+      if !proper_weights?
+        argument_error
+      end
+    end
+
+    def proper_weights?
+      @weights.inject(0) {|sum, v| sum + v} == 1.0
+    end
+
+    def probability_density_function(x)
+      pdf = 0.0
+      @models.each_with_index do |model, mi|
+        pdf += model.pdf(x) * @weights[mi]
+      end
+      pdf
+    end
+
+    def clear_temp_weight_per_datum!
+      @temp_weight_per_datum.each {|w| w.clear}
+    end
+
+    def calculate_posterior_data_array(data_array)
+      posterior_data_array = Array.new(data_array.size, 0.0)
+      @models.each_with_index do |model, mi|
+        data_array.each_with_index do |x, di|
+          posterior_data_array[di] += @weights[mi] * model.pdf(x)
+        end
+      end
+      posterior_data_array
+    end
+
+    def update_temp_weights!(data_array, posterior_data_array)
+      @models.each_with_index do |model, mi|
+        data_array.each_with_index do |x, di|
+          temp_weight_per_datum = @weights[mi] * model.pdf(x) / posterior_data_array[di]
+          temp_weight_per_datum = 0.0 if temp_weight_per_datum.nan?
+          @temp_weight_per_datum[mi] <<  temp_weight_per_datum
+        end
+        @temp_weights[mi] = @temp_weight_per_datum[mi].inject(0.0) {|sum, w| sum + w}
+      end
+    end
+
+    def update_weights!(data_array)
+      (0..(@models.size-1)).each do |mi|
+        @weights[mi] = @temp_weights[mi] / data_array.size
+      end
+    end
+
+    def update_parameters!(data_array)
+      @models.each_with_index do |model, mi|
+        model.update_parameters!(data_array, @temp_weights[mi], @temp_weight_per_datum[mi])
+      end
+      update_weights!(data_array)
+    end
+
+    # output types
+    # :full (default)
+    # :separate_only
+    # :mixture_only
+    def to_gnuplot(type = :full)
+      # output each model (currently assume Gaussian)
+      output = []
+      @models.each_with_index do |model, mi|
+        output << "#{@weights[mi].round(DIGIT)} * #{model.to_gnuplot_with_title(@weights[mi])}"
+      end
+      separate = output.join(", ")
+      # output mixture model (currently assume Gaussian Mixture model)
+      output = []
+      @models.each_with_index do |model, mi|
+        output << "#{@weights[mi].round(DIGIT)} * #{model.to_gnuplot}"
+      end
+      mixture = output.join(" + ")
+      case type
+      when :separate_only
+        return separate
+      when :mixture_only
+        return mixture
+      end
+      "#{separate}, #{mixture}"
+    end
+
+    def debug_output
+      <<-DEBUG_OUT
+  @weights=#{@weights.inspect}
+  @temp_weights=#{@temp_weights.inspect}
+  @models
+   #{@models.inspect}
+      DEBUG_OUT
+    end
+  end
+end