rumale 0.20.3 → 0.22.3

data/lib/rumale/ensemble/stacking_regressor.rb

@@ -0,0 +1,163 @@
+# frozen_string_literal: true
+
+require 'rumale/base/base_estimator'
+require 'rumale/base/regressor'
+
+module Rumale
+  module Ensemble
+    # StackingRegressor is a class that implements regressor with stacking method.
+    #
+    # @example
+    #   estimators = {
+    #     las: Rumale::LinearModel::Lasso.new(reg_param: 1e-2, random_seed: 1),
+    #     mlp: Rumele::NeuralNetwork::MLPRegressor.new(hidden_units: [256], random_seed: 1),
+    #     rnd: Rumale::Ensemble::RandomForestRegressor.new(random_seed: 1)
+    #   }
+    #   meta_estimator = Rumale::LinearModel::Ridge.new(random_seed: 1)
+    #   regressor = Rumale::Ensemble::StackedRegressor.new(
+    #     estimators: estimators, meta_estimator: meta_estimator, random_seed: 1
+    #   )
+    #   regressor.fit(training_samples, traininig_values)
+    #   results = regressor.predict(testing_samples)
+    #
+    # *Reference*
+    # - Zhou, Z-H., "Ensemble Mehotds - Foundations and Algorithms," CRC Press Taylor and Francis Group, Chapman and Hall/CRC, 2012.
+    class StackingRegressor
+      include Base::BaseEstimator
+      include Base::Regressor
+
+      # Return the base regressors.
+      # @return [Hash<Symbol,Regressor>]
+      attr_reader :estimators
+
+      # Return the meta regressor.
+      # @return [Regressor]
+      attr_reader :meta_estimator
+
+      # Create a new regressor with stacking method.
+      #
+      # @param estimators [Hash<Symbol,Regressor>] The base regressors for extracting meta features.
+      # @param meta_estimator [Regressor/Nil] The meta regressor that predicts values.
+      #   If nil is given, Ridge is used.
+      # @param n_splits [Integer] The number of folds for cross validation with k-fold on meta feature extraction in training phase.
+      # @param shuffle [Boolean] The flag indicating whether to shuffle the dataset on cross validation.
+      # @param passthrough [Boolean] The flag indicating whether to concatenate the original features and meta features when training the meta regressor.
+      # @param random_seed [Integer/Nil] The seed value using to initialize the random generator on cross validation.
+      def initialize(estimators:, meta_estimator: nil, n_splits: 5, shuffle: true, passthrough: false, random_seed: nil)
+        check_params_type(Hash, estimators: estimators)
+        check_params_numeric(n_splits: n_splits)
+        check_params_boolean(shuffle: shuffle, passthrough: passthrough)
+        check_params_numeric_or_nil(random_seed: random_seed)
+        @estimators = estimators
+        @meta_estimator = meta_estimator || Rumale::LinearModel::Ridge.new
+        @output_size = nil
+        @params = {}
+        @params[:n_splits] = n_splits
+        @params[:shuffle] = shuffle
+        @params[:passthrough] = passthrough
+        @params[:random_seed] = random_seed || srand
+      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 variables to be used for fitting the model.
+      # @return [StackedRegressor] 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)
+
+        n_samples, n_features = x.shape
+        n_outputs = y.ndim == 1 ? 1 : y.shape[1]
+
+        # training base regressors with all training data.
+        @estimators.each_key { |name| @estimators[name].fit(x, y) }
+
+        # detecting size of output for each base regressor.
+        @output_size = detect_output_size(n_features)
+
+        # extracting meta features with base regressors.
+        n_components = @output_size.values.inject(:+)
+        z = Numo::DFloat.zeros(n_samples, n_components)
+
+        kf = Rumale::ModelSelection::KFold.new(
+          n_splits: @params[:n_splits], shuffle: @params[:shuffle], random_seed: @params[:random_seed]
+        )
+
+        kf.split(x, y).each do |train_ids, valid_ids|
+          x_train = x[train_ids, true]
+          y_train = n_outputs == 1 ? y[train_ids] : y[train_ids, true]
+          x_valid = x[valid_ids, true]
+          f_start = 0
+          @estimators.each_key do |name|
+            est_fold = Marshal.load(Marshal.dump(@estimators[name]))
+            f_last = f_start + @output_size[name]
+            f_position = @output_size[name] == 1 ? f_start : f_start...f_last
+            z[valid_ids, f_position] = est_fold.fit(x_train, y_train).predict(x_valid)
+            f_start = f_last
+          end
+        end
+
+        # concatenating original features.
+        z = Numo::NArray.hstack([z, x]) if @params[:passthrough]
+
+        # training meta regressor.
+        @meta_estimator.fit(z, y)
+
+        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]) The predicted values per sample.
+      def predict(x)
+        x = check_convert_sample_array(x)
+        z = transform(x)
+        @meta_estimator.predict(z)
+      end
+
+      # Transform the given data with the learned model.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to be transformed with the learned model.
+      # @return [Numo::DFloat] (shape: [n_samples, n_components]) The meta features for samples.
+      def transform(x)
+        x = check_convert_sample_array(x)
+        n_samples = x.shape[0]
+        n_components = @output_size.values.inject(:+)
+        z = Numo::DFloat.zeros(n_samples, n_components)
+        f_start = 0
+        @estimators.each_key do |name|
+          f_last = f_start + @output_size[name]
+          f_position = @output_size[name] == 1 ? f_start : f_start...f_last
+          z[true, f_position] = @estimators[name].predict(x)
+          f_start = f_last
+        end
+        z = Numo::NArray.hstack([z, x]) if @params[:passthrough]
+        z
+      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, n_outputs]) The target variables to be used for fitting the model.
+      # @return [Numo::DFloat] (shape: [n_samples, n_components]) The meta features for training data.
+      def fit_transform(x, y)
+        x = check_convert_sample_array(x)
+        y = check_convert_tvalue_array(y)
+        fit(x, y).transform(x)
+      end
+
+      private
+
+      def detect_output_size(n_features)
+        x_dummy = Numo::DFloat.new(2, n_features).rand
+        @estimators.each_key.with_object({}) do |name, obj|
+          output_dummy = @estimators[name].predict(x_dummy)
+          obj[name] = output_dummy.ndim == 1 ? 1 : output_dummy.shape[1]
+        end
+      end
+    end
+  end
+end

data/lib/rumale/feature_extraction/feature_hasher.rb

@@ -67,7 +67,7 @@ module Rumale
       def transform(x)
         raise 'FeatureHasher#transform requires Mmh3 but that is not loaded.' unless enable_mmh3?
 
-        x = [x] unless x.is_a?(Array) # rubocop:disable Style/ArrayCoercion
+        x = [x] unless x.is_a?(Array)
         n_samples = x.size
 
         z = Numo::DFloat.zeros(n_samples, n_features)

data/lib/rumale/feature_extraction/hash_vectorizer.rb

@@ -99,7 +99,7 @@ module Rumale
       # @param x [Array<Hash>] (shape: [n_samples]) The array of hash consisting of feature names and values.
       # @return [Numo::DFloat] (shape: [n_samples, n_features]) The encoded sample array.
       def transform(x)
-        x = [x] unless x.is_a?(Array) # rubocop:disable Style/ArrayCoercion
+        x = [x] unless x.is_a?(Array)
         n_samples = x.size
         n_features = @vocabulary.size
         z = Numo::DFloat.zeros(n_samples, n_features)

data/lib/rumale/kernel_machine/kernel_svc.rb

@@ -11,9 +11,10 @@ module Rumale
     # with stochastic gradient descent (SGD) optimization.
     # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
-    # Rumale::SVM provides kernel support vector classifier based on LIBSVM.
-    # If you prefer execution speed, you should use Rumale::SVM::SVC.
-    # https://github.com/yoshoku/rumale-svm
+    # @note
+    #   Rumale::SVM provides kernel support vector classifier based on LIBSVM.
+    #   If you prefer execution speed, you should use Rumale::SVM::SVC.
+    #   https://github.com/yoshoku/rumale-svm
     #
     # @example
     #   training_kernel_matrix = Rumale::PairwiseMetric::rbf_kernel(training_samples)

data/lib/rumale/linear_model/base_sgd.rb

@@ -171,7 +171,7 @@ module Rumale
         @params[:fit_bias] = true
         @params[:reg_param] = 0.0
         @params[:l1_ratio] = 0.0
-        @params[:max_iter] = 200
+        @params[:max_iter] = 1000
         @params[:batch_size] = 50
         @params[:tol] = 0.0001
         @params[:verbose] = false

data/lib/rumale/linear_model/elastic_net.rb

@@ -10,7 +10,7 @@ module Rumale
     #
     # @example
     #   estimator =
-    #     Rumale::LinearModel::ElasticNet.new(reg_param: 0.1, l1_ratio: 0.5, max_iter: 200, batch_size: 50, random_seed: 1)
+    #     Rumale::LinearModel::ElasticNet.new(reg_param: 0.1, l1_ratio: 0.5, max_iter: 1000, batch_size: 50, random_seed: 1)
     #   estimator.fit(training_samples, traininig_values)
     #   results = estimator.predict(testing_samples)
     #
@@ -59,7 +59,7 @@ module Rumale
       # @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, 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,
                      n_jobs: nil, verbose: false, random_seed: nil)
         check_params_numeric(learning_rate: learning_rate, momentum: momentum,
                              reg_param: reg_param, l1_ratio: l1_ratio, bias_scale: bias_scale,
@@ -81,7 +81,7 @@ 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 [ElasticNet] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)

data/lib/rumale/linear_model/lasso.rb

@@ -10,7 +10,7 @@ module Rumale
     #
     # @example
     #   estimator =
-    #     Rumale::LinearModel::Lasso.new(reg_param: 0.1, max_iter: 500, batch_size: 20, random_seed: 1)
+    #     Rumale::LinearModel::Lasso.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)
     #
@@ -55,7 +55,7 @@ module Rumale
       # @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,
                      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,
@@ -77,7 +77,7 @@ 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 [Lasso] The learned regressor itself.
       def fit(x, y)
         x = check_convert_sample_array(x)

data/lib/rumale/linear_model/linear_regression.rb

@@ -6,11 +6,12 @@ require 'rumale/base/regressor'
 module Rumale
   module LinearModel
     # LinearRegression is a class that implements ordinary least square linear 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::LinearRegression.new(max_iter: 500, batch_size: 20, random_seed: 1)
+    #     Rumale::LinearModel::LinearRegression.new(max_iter: 1000, batch_size: 20, random_seed: 1)
     #   estimator.fit(training_samples, traininig_values)
     #   results = estimator.predict(testing_samples)
     #
@@ -41,34 +42,35 @@ 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 '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 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.
+      #   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,
-                     fit_bias: true, bias_scale: 1.0, max_iter: 200, batch_size: 50, tol: 1e-4,
+                     fit_bias: true, bias_scale: 1.0, 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,
@@ -80,9 +82,9 @@ 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[:learning_rate]
         @params[:random_seed] ||= srand
@@ -95,15 +97,17 @@ 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 [LinearRegression] 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
@@ -124,24 +128,46 @@ module Rumale
 
       def fit_svd(x, y)
         x = expand_feature(x) if fit_bias?
-
         w = Numo::Linalg.pinv(x, driver: 'svd').dot(y)
+        @weight_vec, @bias_term = single_target?(y) ? split_weight(w) : split_weight_mult(w)
+      end
 
-        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])
+      def fit_lbfgs(x, y)
+        fnc = proc do |w, x, y| # 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)
+          gradient = 2.fdiv(n_samples) * d.transpose.dot(x)
+          [loss, gradient.flatten.dup]
         end
-      end
 
-      def fit_sgd(x, y)
-        n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
+        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],
+          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?
@@ -150,20 +176,23 @@ 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 fit_bias?
-        @params[:fit_bias] == true
+      def single_target?(y)
+        y.ndim == 1
       end
 
-      def load_linalg?
-        return false if defined?(Numo::Linalg).nil?
-        return false if Numo::Linalg::VERSION < '0.1.4'
+      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/logistic_regression.rb

@@ -1,21 +1,24 @@
 # frozen_string_literal: true
 
-require 'rumale/linear_model/base_sgd'
+require 'lbfgsb'
 require 'rumale/base/classifier'
+require 'rumale/linear_model/base_sgd'
+require 'rumale/preprocessing/label_binarizer'
 
 module Rumale
   module LinearModel
-    # LogisticRegression is a class that implements Logistic Regression
-    # with stochastic gradient descent optimization.
-    # For multiclass classification problem, it uses one-vs-the-rest strategy.
+    # LogisticRegression is a class that implements Logistic Regression.
+    # In multiclass classification problem, it uses one-vs-the-rest strategy for the sgd solver
+    # and multinomial logistic regression for the lbfgs solver.
     #
-    # Rumale::SVM provides Logistic Regression based on LIBLINEAR.
-    # If you prefer execution speed, you should use Rumale::SVM::LogisticRegression.
-    # https://github.com/yoshoku/rumale-svm
+    # @note
+    #   Rumale::SVM provides Logistic Regression based on LIBLINEAR.
+    #   If you prefer execution speed, you should use Rumale::SVM::LogisticRegression.
+    #   https://github.com/yoshoku/rumale-svm
     #
     # @example
     #   estimator =
-    #     Rumale::LinearModel::LogisticRegression.new(reg_param: 1.0, max_iter: 200, batch_size: 50, random_seed: 1)
+    #     Rumale::LinearModel::LogisticRegression.new(reg_param: 1.0, random_seed: 1)
     #   estimator.fit(training_samples, traininig_labels)
     #   results = estimator.predict(testing_samples)
     #
@@ -42,19 +45,24 @@ module Rumale
       # @return [Random]
       attr_reader :rng
 
-      # Create a new classifier with Logisitc Regression by the SGD optimization.
+      # Create a new classifier with Logisitc Regression.
       #
       # @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 = 'lbfgs', 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 = 'lbfgs', this parameter is ignored.
       # @param momentum [Float] The momentum factor.
+      #   If solver = 'lbfgs', this parameter is ignored.
       # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
+      #   If solver = 'lbfgs', only 'l2' can be selected for this parameter.
       # @param l1_ratio [Float] The elastic-net type regularization mixing parameter.
       #   If penalty set to 'l2' or 'l1', this parameter is ignored.
       #   If l1_ratio = 1, the regularization is similar to Lasso.
       #   If l1_ratio = 0, the regularization is similar to Ridge.
       #   If 0 < l1_ratio < 1, the regularization is a combination of L1 and L2.
+      #   If solver = 'lbfgs', 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.
@@ -62,28 +70,38 @@ module Rumale
       # @param max_iter [Integer] The maximum number of epochs that indicates
       #   how many times the whole data is given to the training process.
       # @param batch_size [Integer] The size of the mini batches.
+      #   If solver = 'lbfgs', this parameter is ignored.
       # @param tol [Float] The tolerance of loss for terminating optimization.
+      #   If solver = 'lbfgs', this value is given as tol / Lbfgsb::DBL_EPSILON to the factr argument of Lbfgsb.minimize method.
+      # @param solver [String] The algorithm for optimization. ('lbfgs' or 'sgd').
+      #   'lbfgs' uses the L-BFGS with lbfgs.rb gem.
+      #   'sgd' uses the stochastic gradient descent optimization.
       # @param n_jobs [Integer] The number of jobs for running the fit and predict methods in parallel.
       #   If nil is given, the methods do 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.
+      #   This parameter is ignored if the Parallel gem is not loaded or the solver is 'lbfgs'.
       # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
+      #   If solver = 'lbfgs' and 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(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,
+                     solver: 'lbfgs',
                      n_jobs: nil, verbose: false, random_seed: nil)
         check_params_numeric(learning_rate: learning_rate, momentum: momentum,
                              reg_param: reg_param, l1_ratio: l1_ratio, bias_scale: bias_scale,
                              max_iter: max_iter, batch_size: batch_size, tol: tol)
         check_params_boolean(fit_bias: fit_bias, verbose: verbose)
-        check_params_string(penalty: penalty)
+        check_params_string(solver: solver, penalty: penalty)
         check_params_numeric_or_nil(decay: decay, n_jobs: n_jobs, random_seed: random_seed)
         check_params_positive(learning_rate: learning_rate, reg_param: reg_param,
                               bias_scale: bias_scale, max_iter: max_iter, batch_size: batch_size)
+        raise ArgumentError, "The 'lbfgs' solver supports only 'l2' penalties." if solver == 'lbfgs' && penalty != 'l2'
+
         super()
         @params.merge!(method(:initialize).parameters.map { |_t, arg| [arg, binding.local_variable_get(arg)] }.to_h)
+        @params[:solver] = solver == 'sgd' ? 'sgd' : 'lbfgs'
         @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
         @params[:random_seed] ||= srand
         @rng = Random.new(@params[:random_seed])
@@ -105,30 +123,10 @@ module Rumale
         check_sample_label_size(x, y)
 
         @classes = Numo::Int32[*y.to_a.uniq.sort]
-
-        if multiclass_problem?
-          n_classes = @classes.size
-          n_features = x.shape[1]
-          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
-          @bias_term = Numo::DFloat.zeros(n_classes)
-          if enable_parallel?
-            # :nocov:
-            models = parallel_map(n_classes) do |n|
-              bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
-              partial_fit(x, bin_y)
-            end
-            # :nocov:
-            n_classes.times { |n| @weight_vec[n, true], @bias_term[n] = models[n] }
-          else
-            n_classes.times do |n|
-              bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
-              @weight_vec[n, true], @bias_term[n] = partial_fit(x, bin_y)
-            end
-          end
+        if @params[:solver] == 'sgd'
+          fit_sgd(x, y)
         else
-          negative_label = @classes[0]
-          bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
-          @weight_vec, @bias_term = partial_fit(x, bin_y)
+          fit_lbfgs(x, y)
         end
 
         self
@@ -182,6 +180,96 @@ module Rumale
       def multiclass_problem?
         @classes.size > 2
       end
+
+      def fit_lbfgs(base_x, base_y) # rubocop:disable Metrics/AbcSize, Metrics/MethodLength
+        if multiclass_problem?
+          fnc = proc do |w, x, y, a|
+            n_features = x.shape[1]
+            n_classes = y.shape[1]
+            z = x.dot(w.reshape(n_classes, n_features).transpose)
+            # logsumexp and softmax
+            z_max = z.max(-1).expand_dims(-1).dup
+            z_max[~z_max.isfinite] = 0.0
+            lgsexp = Numo::NMath.log(Numo::NMath.exp(z - z_max).sum(-1)).expand_dims(-1) + z_max
+            t = z - lgsexp
+            sftmax = Numo::NMath.exp(t)
+            # loss and gradient
+            loss = -(y * t).sum + 0.5 * a * w.dot(w)
+            grad = (sftmax - y).transpose.dot(x).flatten.dup + a * w
+            [loss, grad]
+          end
+
+          base_x = expand_feature(base_x) if fit_bias?
+          encoder = Rumale::Preprocessing::LabelBinarizer.new
+          onehot_y = encoder.fit_transform(base_y)
+          n_classes = @classes.size
+          n_features = base_x.shape[1]
+          w_init = Numo::DFloat.zeros(n_classes * n_features)
+
+          verbose = @params[:verbose] ? 1 : -1
+          res = Lbfgsb.minimize(
+            fnc: fnc, jcb: true, x_init: w_init, args: [base_x, onehot_y, @params[:reg_param]],
+            maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: verbose
+          )
+
+          if fit_bias?
+            weight = res[:x].reshape(n_classes, n_features)
+            @weight_vec = weight[true, 0...-1].dup
+            @bias_term = weight[true, -1].dup
+          else
+            @weight_vec = res[:x].reshape(n_classes, n_features)
+            @bias_term = Numo::DFloat.zeros(n_classes)
+          end
+        else
+          fnc = proc do |w, x, y, a|
+            z = 1 + Numo::NMath.exp(-y * x.dot(w))
+            loss = Numo::NMath.log(z).sum + 0.5 * a * w.dot(w)
+            grad = (y / z - y).dot(x) + a * w
+            [loss, grad]
+          end
+
+          base_x = expand_feature(base_x) if fit_bias?
+          negative_label = @classes[0]
+          bin_y = Numo::Int32.cast(base_y.ne(negative_label)) * 2 - 1
+          n_features = base_x.shape[1]
+          w_init = Numo::DFloat.zeros(n_features)
+
+          verbose = @params[:verbose] ? 1 : -1
+          res = Lbfgsb.minimize(
+            fnc: fnc, jcb: true, x_init: w_init, args: [base_x, bin_y, @params[:reg_param]],
+            maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: verbose
+          )
+
+          @weight_vec, @bias_term = split_weight(res[:x])
+        end
+      end
+
+      def fit_sgd(x, y)
+        if multiclass_problem?
+          n_classes = @classes.size
+          n_features = x.shape[1]
+          @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
+          @bias_term = Numo::DFloat.zeros(n_classes)
+          if enable_parallel?
+            # :nocov:
+            models = parallel_map(n_classes) do |n|
+              bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
+              partial_fit(x, bin_y)
+            end
+            # :nocov:
+            n_classes.times { |n| @weight_vec[n, true], @bias_term[n] = models[n] }
+          else
+            n_classes.times do |n|
+              bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
+              @weight_vec[n, true], @bias_term[n] = partial_fit(x, bin_y)
+            end
+          end
+        else
+          negative_label = @classes[0]
+          bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
+          @weight_vec, @bias_term = partial_fit(x, bin_y)
+        end
+      end
     end
   end
 end