rumale 0.20.3 → 0.22.3

data/lib/rumale/linear_model/nnls.rb

@@ -0,0 +1,137 @@
+# frozen_string_literal: true
+
+require 'lbfgsb'
+
+require 'rumale/base/base_estimator'
+require 'rumale/base/regressor'
+
+module Rumale
+  module LinearModel
+    # NNLS is a class that implements non-negative least squares regression.
+    # NNLS solves least squares problem under non-negative constraints on the coefficient using L-BFGS-B method.
+    #
+    # @example
+    #   estimator = Rumale::LinearModel::NNLS.new(reg_param: 0.01, random_seed: 1)
+    #   estimator.fit(training_samples, traininig_values)
+    #   results = estimator.predict(testing_samples)
+    #
+    class NNLS
+      include Base::BaseEstimator
+      include Base::Regressor
+
+      # Return the weight vector.
+      # @return [Numo::DFloat] (shape: [n_outputs, n_features])
+      attr_reader :weight_vec
+
+      # Return the bias term (a.k.a. intercept).
+      # @return [Numo::DFloat] (shape: [n_outputs])
+      attr_reader :bias_term
+
+      # Returns the number of iterations when converged.
+      # @return [Integer]
+      attr_reader :n_iter
+
+      # Return the random generator for initializing weight.
+      # @return [Random]
+      attr_reader :rng
+
+      # Create a new regressor with non-negative least squares method.
+      #
+      # @param reg_param [Float] The regularization parameter for L2 regularization term.
+      # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
+      # @param bias_scale [Float] The scale of the bias term.
+      # @param max_iter [Integer] The maximum number of epochs that indicates
+      #   how many times the whole data is given to the training process.
+      # @param tol [Float] The tolerance of loss for terminating optimization.
+      #   If solver = 'svd', this parameter is ignored.
+      # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
+      # @param random_seed [Integer] The seed value using to initialize the random generator.
+      def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0,
+                     max_iter: 1000, tol: 1e-4, verbose: false, random_seed: nil)
+        check_params_numeric(reg_param: reg_param, bias_scale: bias_scale, max_iter: max_iter, tol: tol)
+        check_params_boolean(fit_bias: fit_bias, verbose: verbose)
+        check_params_numeric_or_nil(random_seed: random_seed)
+        check_params_positive(reg_param: reg_param, max_iter: max_iter)
+        @params = method(:initialize).parameters.each_with_object({}) { |(_, prm), obj| obj[prm] = binding.local_variable_get(prm) }
+        @params[:random_seed] ||= srand
+        @n_iter = nil
+        @weight_vec = nil
+        @bias_term = nil
+        @rng = Random.new(@params[:random_seed])
+      end
+
+      # Fit the model with given training data.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
+      # @return [NonneagtiveLeastSquare] The learned regressor itself.
+      def fit(x, y)
+        x = check_convert_sample_array(x)
+        y = check_convert_tvalue_array(y)
+        check_sample_tvalue_size(x, y)
+
+        x = expand_feature(x) if fit_bias?
+
+        n_features = x.shape[1]
+        n_outputs = single_target?(y) ? 1 : y.shape[1]
+
+        w_init = Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
+        w_init[w_init.lt(0)] = 0
+        bounds = Numo::DFloat.zeros(n_outputs * n_features, 2)
+        bounds.shape[0].times { |n| bounds[n, 1] = Float::INFINITY }
+
+        res = Lbfgsb.minimize(
+          fnc: method(:nnls_fnc), jcb: true, x_init: w_init, args: [x, y, @params[:reg_param]], bounds: bounds,
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: @params[:verbose] ? 1 : -1
+        )
+
+        @n_iter = res[:n_iter]
+        w = single_target?(y) ? res[:x] : res[:x].reshape(n_outputs, n_features).transpose
+
+        if fit_bias?
+          @weight_vec = single_target?(y) ? w[0...-1].dup : w[0...-1, true].dup
+          @bias_term = single_target?(y) ? w[-1] : w[-1, true].dup
+        else
+          @weight_vec = w.dup
+          @bias_term = single_target?(y) ? 0 : Numo::DFloat.zeros(y.shape[1])
+        end
+
+        self
+      end
+
+      # Predict values for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the values.
+      # @return [Numo::DFloat] (shape: [n_samples, n_outputs]) Predicted values per sample.
+      def predict(x)
+        x = check_convert_sample_array(x)
+        x.dot(@weight_vec.transpose) + @bias_term
+      end
+
+      private
+
+      def nnls_fnc(w, x, y, alpha)
+        n_samples, n_features = x.shape
+        w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
+        z = x.dot(w.transpose)
+        d = z - y
+        loss = (d**2).sum.fdiv(n_samples) + alpha * (w * w).sum
+        gradient = 2.fdiv(n_samples) * d.transpose.dot(x) + 2.0 * alpha * w
+        [loss, gradient.flatten.dup]
+      end
+
+      def expand_feature(x)
+        n_samples = x.shape[0]
+        Numo::NArray.hstack([x, Numo::DFloat.ones([n_samples, 1]) * @params[:bias_scale]])
+      end
+
+      def fit_bias?
+        @params[:fit_bias] == true
+      end
+
+      def single_target?(y)
+        y.ndim == 1
+      end
+    end
+  end
+end

data/lib/rumale/linear_model/ridge.rb

@@ -1,16 +1,19 @@
 # frozen_string_literal: true
 
+require 'lbfgsb'
+
 require 'rumale/linear_model/base_sgd'
 require 'rumale/base/regressor'
 
 module Rumale
   module LinearModel
     # Ridge is a class that implements Ridge Regression
-    # with stochastic gradient descent (SGD) optimization or singular value decomposition (SVD).
+    # with stochastic gradient descent (SGD) optimization,
+    # singular value decomposition (SVD), or L-BFGS optimization.
     #
     # @example
     #   estimator =
-    #     Rumale::LinearModel::Ridge.new(reg_param: 0.1, max_iter: 500, batch_size: 20, random_seed: 1)
+    #     Rumale::LinearModel::Ridge.new(reg_param: 0.1, max_iter: 1000, batch_size: 20, random_seed: 1)
     #   estimator.fit(training_samples, traininig_values)
     #   results = estimator.predict(testing_samples)
     #
@@ -41,36 +44,37 @@ module Rumale
       #
       # @param learning_rate [Float] The initial value of learning rate.
       #   The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
       #   If nil is given, the decay sets to 'reg_param * learning_rate'.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param momentum [Float] The momentum factor.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param reg_param [Float] The regularization parameter.
       # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
       # @param bias_scale [Float] The scale of the bias term.
       # @param max_iter [Integer] The maximum number of epochs that indicates
       #   how many times the whole data is given to the training process.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is 'svd', this parameter is ignored.
       # @param batch_size [Integer] The size of the mini batches.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is not 'sgd', this parameter is ignored.
       # @param tol [Float] The tolerance of loss for terminating optimization.
-      #   If solver = 'svd', this parameter is ignored.
-      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd' or 'svd').
+      #   If solver is 'svd', this parameter is ignored.
+      # @param solver [String] The algorithm to calculate weights. ('auto', 'sgd', 'svd', or 'lbfgs').
       #   'auto' chooses the 'svd' solver if Numo::Linalg is loaded. Otherwise, it chooses the 'sgd' solver.
       #   'sgd' uses the stochastic gradient descent optimization.
       #   'svd' performs singular value decomposition of samples.
+      #   'lbfgs' uses the L-BFGS method for optimization.
       # @param n_jobs [Integer] The number of jobs for running the fit method in parallel.
       #   If nil is given, the method does not execute in parallel.
       #   If zero or less is given, it becomes equal to the number of processors.
-      #   This parameter is ignored if the Parallel gem is not loaded or the solver is 'svd'.
+      #   This parameter is ignored if the Parallel gem is not loaded or solver is not 'sgd'.
       # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
-      #   If solver = 'svd', this parameter is ignored.
+      #   If solver is 'svd', this parameter is ignored.
       # @param random_seed [Integer] The seed value using to initialize the random generator.
       def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
                      reg_param: 1.0, fit_bias: true, bias_scale: 1.0,
-                     max_iter: 200, batch_size: 50, tol: 1e-4,
+                     max_iter: 1000, batch_size: 50, tol: 1e-4,
                      solver: 'auto',
                      n_jobs: nil, verbose: false, random_seed: nil)
         check_params_numeric(learning_rate: learning_rate, momentum: momentum,
@@ -83,9 +87,9 @@ module Rumale
         super()
         @params.merge!(method(:initialize).parameters.map { |_t, arg| [arg, binding.local_variable_get(arg)] }.to_h)
         @params[:solver] = if solver == 'auto'
-                             load_linalg? ? 'svd' : 'sgd'
+                             enable_linalg?(warning: false) ? 'svd' : 'sgd'
                            else
-                             solver != 'svd' ? 'sgd' : 'svd'
+                             solver.match?(/^svd$|^sgd$|^lbfgs$/) ? solver : 'sgd'
                            end
         @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
         @params[:random_seed] ||= srand
@@ -99,15 +103,17 @@ module Rumale
       # Fit the model with given training data.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
-      # @param y [Numo::Int32] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
       # @return [Ridge] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)
         y = check_convert_tvalue_array(y)
         check_sample_tvalue_size(x, y)
 
-        if @params[:solver] == 'svd' && enable_linalg?
+        if @params[:solver] == 'svd' && enable_linalg?(warning: false)
           fit_svd(x, y)
+        elsif @params[:solver] == 'lbfgs'
+          fit_lbfgs(x, y)
         else
           fit_sgd(x, y)
         end
@@ -127,27 +133,51 @@ module Rumale
       private
 
       def fit_svd(x, y)
-        samples = @params[:fit_bias] ? expand_feature(x) : x
+        x = expand_feature(x) if fit_bias?
 
-        s, u, vt = Numo::Linalg.svd(samples, driver: 'sdd', job: 'S')
+        s, u, vt = Numo::Linalg.svd(x, driver: 'sdd', job: 'S')
         d = (s / (s**2 + @params[:reg_param])).diag
         w = vt.transpose.dot(d).dot(u.transpose).dot(y)
 
-        is_single_target_vals = y.shape[1].nil?
-        if @params[:fit_bias]
-          @weight_vec = is_single_target_vals ? w[0...-1].dup : w[0...-1, true].dup
-          @bias_term = is_single_target_vals ? w[-1] : w[-1, true].dup
-        else
-          @weight_vec = w.dup
-          @bias_term = is_single_target_vals ? 0 : Numo::DFloat.zeros(y.shape[1])
-        end
+        @weight_vec, @bias_term = single_target?(y) ? split_weight(w) : split_weight_mult(w)
       end
 
-      def fit_sgd(x, y)
-        n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
+      def fit_lbfgs(x, y)
+        fnc = proc do |w, x, y, a| # rubocop:disable Lint/ShadowingOuterLocalVariable
+          n_samples, n_features = x.shape
+          w = w.reshape(y.shape[1], n_features) unless y.shape[1].nil?
+          z = x.dot(w.transpose)
+          d = z - y
+          loss = (d**2).sum.fdiv(n_samples) + a * (w * w).sum
+          gradient = 2.fdiv(n_samples) * d.transpose.dot(x) + 2.0 * a * w
+          [loss, gradient.flatten.dup]
+        end
+
+        x = expand_feature(x) if fit_bias?
+
         n_features = x.shape[1]
+        n_outputs = single_target?(y) ? 1 : y.shape[1]
+
+        res = Lbfgsb.minimize(
+          fnc: fnc, jcb: true, x_init: init_weight(n_features, n_outputs), args: [x, y, @params[:reg_param]],
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
+          verbose: @params[:verbose] ? 1 : -1
+        )
+
+        @weight_vec, @bias_term =
+          if single_target?(y)
+            split_weight(res[:x])
+          else
+            split_weight_mult(res[:x].reshape(n_outputs, n_features).transpose)
+          end
+      end
 
-        if n_outputs > 1
+      def fit_sgd(x, y)
+        if single_target?(y)
+          @weight_vec, @bias_term = partial_fit(x, y)
+        else
+          n_outputs = y.shape[1]
+          n_features = x.shape[1]
           @weight_vec = Numo::DFloat.zeros(n_outputs, n_features)
           @bias_term = Numo::DFloat.zeros(n_outputs)
           if enable_parallel?
@@ -156,16 +186,23 @@ module Rumale
           else
             n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = partial_fit(x, y[true, n]) }
           end
-        else
-          @weight_vec, @bias_term = partial_fit(x, y)
         end
       end
 
-      def load_linalg?
-        return false if defined?(Numo::Linalg).nil?
-        return false if Numo::Linalg::VERSION < '0.1.4'
+      def single_target?(y)
+        y.ndim == 1
+      end
+
+      def init_weight(n_features, n_outputs)
+        Rumale::Utils.rand_normal([n_outputs, n_features], @rng.dup).flatten.dup
+      end
 
-        true
+      def split_weight_mult(w)
+        if fit_bias?
+          [w[0...-1, true].dup, w[-1, true].dup]
+        else
+          [w.dup, Numo::DFloat.zeros(w.shape[1])]
+        end
       end
     end
   end

data/lib/rumale/linear_model/svc.rb

@@ -11,13 +11,14 @@ module Rumale
     # with stochastic gradient descent optimization.
     # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
-    # Rumale::SVM provides linear support vector classifier based on LIBLINEAR.
-    # If you prefer execution speed, you should use Rumale::SVM::LinearSVC.
-    # https://github.com/yoshoku/rumale-svm
+    # @note
+    #   Rumale::SVM provides linear support vector classifier based on LIBLINEAR.
+    #   If you prefer execution speed, you should use Rumale::SVM::LinearSVC.
+    #   https://github.com/yoshoku/rumale-svm
     #
     # @example
     #   estimator =
-    #     Rumale::LinearModel::SVC.new(reg_param: 1.0, max_iter: 200, batch_size: 50, random_seed: 1)
+    #     Rumale::LinearModel::SVC.new(reg_param: 1.0, max_iter: 1000, batch_size: 50, random_seed: 1)
     #   estimator.fit(training_samples, traininig_labels)
     #   results = estimator.predict(testing_samples)
     #
@@ -74,7 +75,7 @@ module Rumale
       def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
                      penalty: 'l2', reg_param: 1.0, l1_ratio: 0.5,
                      fit_bias: true, bias_scale: 1.0,
-                     max_iter: 200, batch_size: 50, tol: 1e-4,
+                     max_iter: 1000, batch_size: 50, tol: 1e-4,
                      probability: false,
                      n_jobs: nil, verbose: false, random_seed: nil)
         check_params_numeric(learning_rate: learning_rate, momentum: momentum,

data/lib/rumale/linear_model/svr.rb

@@ -8,13 +8,14 @@ module Rumale
     # SVR is a class that implements Support Vector Regressor
     # with stochastic gradient descent optimization.
     #
-    # Rumale::SVM provides linear and kernel support vector regressor based on LIBLINEAR and LIBSVM.
-    # If you prefer execution speed, you should use Rumale::SVM::LinearSVR.
-    # https://github.com/yoshoku/rumale-svm
+    # @note
+    #   Rumale::SVM provides linear and kernel support vector regressor based on LIBLINEAR and LIBSVM.
+    #   If you prefer execution speed, you should use Rumale::SVM::LinearSVR.
+    #   https://github.com/yoshoku/rumale-svm
     #
     # @example
     #   estimator =
-    #     Rumale::LinearModel::SVR.new(reg_param: 1.0, epsilon: 0.1, max_iter: 200, batch_size: 50, random_seed: 1)
+    #     Rumale::LinearModel::SVR.new(reg_param: 1.0, epsilon: 0.1, max_iter: 1000, batch_size: 50, random_seed: 1)
     #   estimator.fit(training_samples, traininig_target_values)
     #   results = estimator.predict(testing_samples)
     #
@@ -68,7 +69,7 @@ module Rumale
                      penalty: 'l2', reg_param: 1.0, l1_ratio: 0.5,
                      fit_bias: true, bias_scale: 1.0,
                      epsilon: 0.1,
-                     max_iter: 200, batch_size: 50, tol: 1e-4,
+                     max_iter: 1000, batch_size: 50, tol: 1e-4,
                      n_jobs: nil, verbose: false, random_seed: nil)
         check_params_numeric(learning_rate: learning_rate, momentum: momentum,
                              reg_param: reg_param, bias_scale: bias_scale, epsilon: epsilon,

data/lib/rumale/metric_learning/mlkr.rb

@@ -0,0 +1,161 @@
+# frozen_string_literal: true
+
+require 'rumale/base/base_estimator'
+require 'rumale/base/transformer'
+require 'rumale/decomposition/pca'
+require 'rumale/pairwise_metric'
+require 'rumale/utils'
+require 'lbfgsb'
+
+module Rumale
+  module MetricLearning
+    # MLKR is a class that implements Metric Learning for Kernel Regression.
+    #
+    # @example
+    #   transformer = Rumale::MetricLearning::MLKR.new
+    #   transformer.fit(training_samples, traininig_target_values)
+    #   low_samples = transformer.transform(testing_samples)
+    #
+    # *Reference*
+    # - Weinberger, K. Q. and Tesauro, G., "Metric Learning for Kernel Regression," Proc. AISTATS'07, pp. 612--629, 2007.
+    class MLKR
+      include Base::BaseEstimator
+      include Base::Transformer
+
+      # Returns the metric components.
+      # @return [Numo::DFloat] (shape: [n_components, n_features])
+      attr_reader :components
+
+      # Return the number of iterations run for optimization
+      # @return [Integer]
+      attr_reader :n_iter
+
+      # Return the random generator.
+      # @return [Random]
+      attr_reader :rng
+
+      # Create a new transformer with MLKR.
+      #
+      # @param n_components [Integer] The number of components.
+      # @param init [String] The initialization method for components ('random' or 'pca').
+      # @param max_iter [Integer] The maximum number of iterations.
+      # @param tol [Float] The tolerance of termination criterion.
+      #   This value is given as tol / Lbfgsb::DBL_EPSILON to the factr argument of Lbfgsb.minimize method.
+      # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
+      #   If true is given, 'iterate.dat' file is generated by lbfgsb.rb.
+      # @param random_seed [Integer] The seed value using to initialize the random generator.
+      def initialize(n_components: nil, init: 'random', max_iter: 100, tol: 1e-6, verbose: false, random_seed: nil)
+        check_params_numeric_or_nil(n_components: n_components, random_seed: random_seed)
+        check_params_numeric(max_iter: max_iter, tol: tol)
+        check_params_string(init: init)
+        check_params_boolean(verbose: verbose)
+        @params = {}
+        @params[:n_components] = n_components
+        @params[:init] = init
+        @params[:max_iter] = max_iter
+        @params[:tol] = tol
+        @params[:verbose] = verbose
+        @params[:random_seed] = random_seed
+        @params[:random_seed] ||= srand
+        @components = nil
+        @n_iter = nil
+        @rng = Random.new(@params[:random_seed])
+      end
+
+      # Fit the model with given training data.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples]) The target values to be used for fitting the model.
+      # @return [MLKR] The learned classifier itself.
+      def fit(x, y)
+        x = check_convert_sample_array(x)
+        y = check_convert_tvalue_array(y)
+        check_sample_tvalue_size(x, y)
+        n_features = x.shape[1]
+        n_components = if @params[:n_components].nil?
+                         n_features
+                       else
+                         [n_features, @params[:n_components]].min
+                       end
+        @components, @n_iter = optimize_components(x, y, n_features, n_components)
+        @prototypes = x.dot(@components.transpose)
+        @values = y
+        self
+      end
+
+      # Fit the model with training data, and then transform them with the learned model.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
+      # @param y [Numo::DFloat] (shape: [n_samples]) The target values to be used for fitting the model.
+      # @return [Numo::DFloat] (shape: [n_samples, n_components]) The transformed data
+      def fit_transform(x, y)
+        x = check_convert_sample_array(x)
+        y = check_convert_tvalue_array(y)
+        check_sample_tvalue_size(x, y)
+        fit(x, y).transform(x)
+      end
+
+      # Transform the given data with the learned model.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The data to be transformed with the learned model.
+      # @return [Numo::DFloat] (shape: [n_samples, n_components]) The transformed data.
+      def transform(x)
+        x = check_convert_sample_array(x)
+        x.dot(@components.transpose)
+      end
+
+      private
+
+      def init_components(x, n_features, n_components)
+        if @params[:init] == 'pca'
+          pca = Rumale::Decomposition::PCA.new(n_components: n_components)
+          pca.fit(x).components.flatten.dup
+        else
+          Rumale::Utils.rand_normal([n_features, n_components], @rng.dup).flatten.dup
+        end
+      end
+
+      def optimize_components(x, y, n_features, n_components)
+        # initialize components.
+        comp_init = init_components(x, n_features, n_components)
+        # initialize optimization results.
+        res = {}
+        res[:x] = comp_init
+        res[:n_iter] = 0
+        # perform optimization.
+        verbose = @params[:verbose] ? 1 : -1
+        res = Lbfgsb.minimize(
+          fnc: method(:mlkr_fnc), jcb: true, x_init: comp_init, args: [x, y],
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: verbose
+        )
+        # return the results.
+        n_iter = res[:n_iter]
+        comps = n_components == 1 ? res[:x].dup : res[:x].reshape(n_components, n_features)
+        [comps, n_iter]
+      end
+
+      def mlkr_fnc(w, x, y)
+        # initialize some variables.
+        n_features = x.shape[1]
+        n_components = w.size / n_features
+        # projection.
+        w = w.reshape(n_components, n_features)
+        z = x.dot(w.transpose)
+        # predict values.
+        kernel_mat = Numo::NMath.exp(-Rumale::PairwiseMetric.squared_error(z))
+        kernel_mat[kernel_mat.diag_indices] = 0.0
+        norm = kernel_mat.sum(1)
+        norm[norm.eq(0)] = 1
+        y_pred = kernel_mat.dot(y) / norm
+        # calculate loss.
+        y_diff = y_pred - y
+        loss = (y_diff**2).sum
+        # calculate gradient.
+        weight_mat = y_diff * y_diff.expand_dims(1) * kernel_mat
+        weight_mat = weight_mat.sum(0).diag - weight_mat
+        gradient = 8 * z.transpose.dot(weight_mat).dot(x)
+        [loss, gradient.flatten.dup]
+      end
+    end
+  end
+end