rumale 0.20.0 → 0.22.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 358515f8785eb3de2e6571a957ca76cece6b774bb022c1a0951c92d44ab422b4
-  data.tar.gz: '0289b7eb382cd3300845412af0fd43626f4f827bb719083c879b574e3ab37eb0'
+  metadata.gz: 4e2f68b3182ada73537901e7bc74bddd100aff75264f9147c88d8240fb624e29
+  data.tar.gz: e2639a55fc84d1399b925f65b3a56b38f2ae3150dd15ab8556120af28d408cae
 SHA512:
-  metadata.gz: f03fc0f27f99ed4acea3fb7d7bf34017c1dbf923b20dabc9a78d6d44f0b151bc9dc78ba24d122f81607a43fd1852e398a603b75b87656a2f79109f87c0db0d98
-  data.tar.gz: 69f6b8892f6bfb4c43706513245c3fba687dcb6a347c1c5185a70d5e45a024b2848a019bfae48726e1f49212878e8d6d67c811ec5f4a990fdbb3a2841efdfe9b
+  metadata.gz: 91ffcbade578bbb9c6a5d87a54ebd89a2b5990eb70835e7a5549afe78541dbfeafe3af50833725bee751fa89c059484970e5add7ebf8adee3e25bc000fbe3778
+  data.tar.gz: 2ee2b1448a486581ef98561f65bc3446b2e161c89a3a12bd6cd78867350e26151bc0b350bd431902d21f6979493ab2d01a6ee81b55c1099f631aa84c84a704e6

data/.github/workflows/build.yml

@@ -0,0 +1,23 @@
+name: build
+
+on: [push]
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        ruby: [ '2.5', '2.6', '2.7' ]
+    steps:
+      - uses: actions/checkout@v2
+      - name: Install BLAS and LAPACK
+        run: sudo apt-get install -y libopenblas-dev liblapacke-dev
+      - name: Set up Ruby ${{ matrix.ruby }}
+        uses: actions/setup-ruby@v1
+        with:
+          ruby-version: ${{ matrix.ruby }}
+      - name: Build and test with Rake
+        run: |
+          gem install bundler
+          bundle install --jobs 4 --retry 3
+          bundle exec rake

data/.rubocop.yml

@@ -3,6 +3,7 @@ require:
   - rubocop-rspec
 
 AllCops:
+  NewCops: enable
   TargetRubyVersion: 2.5
   DisplayCopNames: true
   DisplayStyleGuide: true
@@ -15,34 +16,16 @@ AllCops:
 Style/Documentation:
   Enabled: false
 
-Style/HashEachMethods:
-  Enabled: true
-
-Style/HashTransformKeys:
-  Enabled: true
-
-Style/HashTransformValues:
-  Enabled: true
-
-Lint/DeprecatedOpenSSLConstant:
-  Enabled: true
-
-Lint/DuplicateElsifCondition:
-  Enabled: true
-
-Lint/MixedRegexpCaptureTypes:
-  Enabled: true
-
-Lint/RaiseException:
-  Enabled: true
-
-Lint/StructNewOverride:
-  Enabled: true
-
 Layout/LineLength:
   Max: 145
   IgnoredPatterns: ['(\A|\s)#']
 
+Lint/ConstantDefinitionInBlock:
+  Enabled: false
+
+Lint/MissingSuper:
+  Enabled: false
+
 Metrics/ModuleLength:
   Max: 200
 
@@ -78,87 +61,21 @@ Naming/MethodParameterName:
 Naming/ConstantName:
   Enabled: false
 
-Style/AccessorGrouping:
-  Enabled: true
-
-Style/ArrayCoercion:
-  Enabled: true
-
-Style/BisectedAttrAccessor:
-  Enabled: true
-
-Style/CaseLikeIf:
-  Enabled: true
-
-Style/ExponentialNotation:
-  Enabled: true
-
 Style/FormatStringToken:
   Enabled: false
 
-Style/HashAsLastArrayItem:
-  Enabled: true
-
-Style/HashLikeCase:
-  Enabled: true
-
 Style/NumericLiterals:
   Enabled: false
 
-Style/RedundantAssignment:
-  Enabled: true
-
-Style/RedundantFetchBlock:
-  Enabled: true
-
-Style/RedundantFileExtensionInRequire:
-  Enabled: true
-
-Style/RedundantRegexpCharacterClass:
-  Enabled: true
-
-Style/RedundantRegexpEscape:
-  Enabled: true
-
-Style/SlicingWithRange:
-  Enabled: true
-
-Layout/EmptyLineAfterGuardClause:
-  Enabled: true
-
-Layout/EmptyLinesAroundAttributeAccessor:
-  Enabled: true
-
-Layout/SpaceAroundMethodCallOperator:
-  Enabled: true
-
-Performance/AncestorsInclude:
-  Enabled: true
-
-Performance/BigDecimalWithNumericArgument:
-  Enabled: true
-
-Performance/RedundantSortBlock:
-  Enabled: true
-
-Performance/RedundantStringChars:
-  Enabled: true
-
-Performance/ReverseFirst:
-  Enabled: true
-
-Performance/SortReverse:
-  Enabled: true
-
-Performance/Squeeze:
-  Enabled: true
-
-Performance/StringInclude:
-  Enabled: true
+Style/StringConcatenation:
+  Enabled: false
 
 RSpec/MultipleExpectations:
   Enabled: false
 
+RSpec/MultipleMemoizedHelpers:
+  Max: 25
+
 RSpec/NestedGroups:
   Max: 4
 
@@ -170,3 +87,6 @@ RSpec/InstanceVariable:
 
 RSpec/LeakyConstantDeclaration:
   Enabled: false
+
+Performance/Sum:
+  Enabled: false

data/CHANGELOG.md

@@ -1,3 +1,31 @@
+# 0.22.0
+## Breaking change
+- Add lbfgsb.rb gem to runtime dependencies. Rumale uses lbfgsb gem for optimization.
+This eliminates the need to require the mopti gem when using [NeighbourhoodComponentAnalysis](https://yoshoku.github.io/rumale/doc/Rumale/MetricLearning/NeighbourhoodComponentAnalysis.html).
+- Add lbfgs solver to [LogisticRegression](https://yoshoku.github.io/rumale/doc/Rumale/LinearModel/LogisticRegression.html) and make it the default solver.
+
+# 0.21.0
+## Breaking change
+- Change the default value of max_iter argument on LinearModel estimators to 1000.
+
+# 0.20.3
+- Fix to use automatic solver of PCA in NeighbourhoodComponentAnalysis.
+- Refactor some codes with Rubocop.
+- Update README.
+
+# 0.20.2
+- Add cross-validator class for time-series data.
+  - [TimeSeriesSplit](https://yoshoku.github.io/rumale/doc/Rumale/ModelSelection/TimeSeriesSplit.html)
+
+# 0.20.1
+- Add cross-validator classes that split data according group labels.
+  - [GroupKFold](https://yoshoku.github.io/rumale/doc/Rumale/ModelSelection/GroupKFold.html)
+  - [GroupShuffleSplit](https://yoshoku.github.io/rumale/doc/Rumale/ModelSelection/GroupShuffleSplit.html)
+- Fix fraction treating of the number of samples on shuffle split cross-validator classes.
+  - [ShuffleSplit](https://yoshoku.github.io/rumale/doc/Rumale/ModelSelection/ShuffleSplit.html)
+  - [StratifiedShuffleSplit](https://yoshoku.github.io/rumale/doc/Rumale/ModelSelection/StratifiedShuffleSplit.html)
+- Refactor some codes with Rubocop.
+
 # 0.20.0
 ## Breaking changes
 - Delete deprecated estimators such as PolynomialModel, Optimizer, and BaseLinearModel.

data/Gemfile

@@ -3,11 +3,13 @@ source 'https://rubygems.org'
 # Specify your gem's dependencies in rumale.gemspec
 gemspec
 
-gem 'coveralls', '~> 0.8'
 gem 'mmh3', '>= 1.0'
-gem 'mopti', '>= 0.1.0'
 gem 'numo-linalg', '>= 0.1.4'
 gem 'parallel', '>= 1.17.0'
 gem 'rake', '~> 12.0'
 gem 'rake-compiler', '~> 1.0'
 gem 'rspec', '~> 3.0'
+gem 'rubocop', '~> 0.91'
+gem 'rubocop-performance', '~> 1.8'
+gem 'rubocop-rspec', '~> 1.43'
+gem 'simplecov', '~> 0.19'

data/README.md

@@ -2,8 +2,7 @@
 
 ![Rumale](https://dl.dropboxusercontent.com/s/joxruk2720ur66o/rumale_header_400.png)
 
-[![Build Status](https://travis-ci.org/yoshoku/rumale.svg?branch=master)](https://travis-ci.org/yoshoku/rumale)
-[![Coverage Status](https://coveralls.io/repos/github/yoshoku/rumale/badge.svg?branch=master)](https://coveralls.io/github/yoshoku/rumale?branch=master)
+[![Build Status](https://github.com/yoshoku/rumale/workflows/build/badge.svg)](https://github.com/yoshoku/rumale/actions?query=workflow%3Abuild)
 [![Gem Version](https://badge.fury.io/rb/rumale.svg)](https://badge.fury.io/rb/rumale)
 [![BSD 2-Clause License](https://img.shields.io/badge/License-BSD%202--Clause-orange.svg)](https://github.com/yoshoku/rumale/blob/master/LICENSE.txt)
 [![Documentation](https://img.shields.io/badge/api-reference-blue.svg)](https://yoshoku.github.io/rumale/doc/)
@@ -228,6 +227,10 @@ When -1 is given to n_jobs parameter, all processors are used.
 estimator = Rumale::Ensemble::RandomForestClassifier.new(n_jobs: -1, random_seed: 1)
 ```
 
+## Novelties
+
+* [Rumale SHOP](https://suzuri.jp/yoshoku)
+
 ## Contributing
 
 Bug reports and pull requests are welcome on GitHub at https://github.com/yoshoku/rumale.

data/lib/rumale.rb

@@ -98,9 +98,12 @@ require 'rumale/preprocessing/ordinal_encoder'
 require 'rumale/preprocessing/binarizer'
 require 'rumale/preprocessing/polynomial_features'
 require 'rumale/model_selection/k_fold'
+require 'rumale/model_selection/group_k_fold'
 require 'rumale/model_selection/stratified_k_fold'
 require 'rumale/model_selection/shuffle_split'
+require 'rumale/model_selection/group_shuffle_split'
 require 'rumale/model_selection/stratified_shuffle_split'
+require 'rumale/model_selection/time_series_split'
 require 'rumale/model_selection/cross_validation'
 require 'rumale/model_selection/grid_search_cv'
 require 'rumale/model_selection/function'

data/lib/rumale/clustering/hdbscan.rb

@@ -136,7 +136,7 @@ module Rumale
         res
       end
 
-      # rubocop:disable Metrics/AbcSize, Metrics/MethodLength, Metrics/PerceivedComplexity
+      # rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity
       def condense_tree(hierarchy, min_cluster_size)
         n_edges = hierarchy.size
         root = 2 * n_edges
@@ -265,7 +265,7 @@ module Rumale
         end
         res
       end
-      # rubocop:enable Metrics/AbcSize, Metrics/MethodLength, Metrics/PerceivedComplexity
+      # rubocop:enable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity
     end
   end
 end

data/lib/rumale/clustering/snn.rb

@@ -51,7 +51,7 @@ module Rumale
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to be used for cluster analysis.
       #   If the metric is 'precomputed', x must be a square distance matrix (shape: [n_samples, n_samples]).
       # @return [Numo::Int32] (shape: [n_samples]) Predicted cluster label per sample.
-      def fit_predict(x)
+      def fit_predict(x) # rubocop:disable Lint/UselessMethodDefinition
         super
       end
 

data/lib/rumale/dataset.rb

@@ -225,7 +225,7 @@ module Rumale
         line = dump_label(label, label_type.to_s)
         ftvec.to_a.each_with_index do |val, n|
           idx = n + (zero_based == false ? 1 : 0)
-          line += format(" %d:#{value_type}", idx, val) if val != 0.0
+          line += format(" %d:#{value_type}", idx, val) if val != 0
         end
         line
       end

data/lib/rumale/decomposition/nmf.rb

@@ -77,7 +77,7 @@ module Rumale
       # @return [Numo::DFloat] (shape: [n_samples, n_components]) The transformed data.
       def transform(x)
         x = check_convert_sample_array(x)
-        partial_fit(x, false)
+        partial_fit(x, update_comps: false)
       end
 
       # Inverse transform the given transformed data with the learned model.
@@ -91,7 +91,7 @@ module Rumale
 
       private
 
-      def partial_fit(x, update_comps = true)
+      def partial_fit(x, update_comps: true)
         # initialize some variables.
         n_samples, n_features = x.shape
         scale = Math.sqrt(x.mean / @params[:n_components])

data/lib/rumale/ensemble/random_forest_classifier.rb

@@ -85,7 +85,7 @@ module Rumale
       # @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]) The labels to be used for fitting the model.
       # @return [RandomForestClassifier] The learned classifier itself.
-      def fit(x, y)
+      def fit(x, y) # rubocop:disable Metrics/AbcSize
         x = check_convert_sample_array(x)
         y = check_convert_label_array(y)
         check_sample_label_size(x, y)

data/lib/rumale/ensemble/random_forest_regressor.rb

@@ -79,7 +79,7 @@ module Rumale
       # @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 [RandomForestRegressor] The learned regressor itself.
-      def fit(x, y)
+      def fit(x, y) # rubocop:disable Metrics/AbcSize
         x = check_convert_sample_array(x)
         y = check_convert_tvalue_array(y)
         check_sample_tvalue_size(x, y)

data/lib/rumale/evaluation_measure/roc_auc.rb

@@ -75,9 +75,12 @@ module Rumale
         false_pos, true_pos, thresholds = binary_roc_curve(y_true, y_score, pos_label)
 
         if true_pos.size.zero? || false_pos[0] != 0 || true_pos[0] != 0
+          # NOTE: Numo::NArray#insert is not a destructive method.
+          # rubocop:disable Style/RedundantSelfAssignment
           true_pos = true_pos.insert(0, 0)
           false_pos = false_pos.insert(0, 0)
           thresholds = thresholds.insert(0, thresholds[0] + 1)
+          # rubocop:enable Style/RedundantSelfAssignment
         end
 
         tpr = true_pos / true_pos[-1].to_f

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)
+        x = [x] unless x.is_a?(Array) # rubocop:disable Style/ArrayCoercion
         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)
+        x = [x] unless x.is_a?(Array) # rubocop:disable Style/ArrayCoercion
         n_samples = x.size
         n_features = @vocabulary.size
         z = Numo::DFloat.zeros(n_samples, n_features)

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,

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,

data/lib/rumale/linear_model/linear_regression.rb

@@ -10,7 +10,7 @@ module Rumale
     #
     # @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)
     #
@@ -68,7 +68,7 @@ module Rumale
       #   If solver = '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,

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)
+        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