rumale 0.22.2 → 0.22.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 703a6895f4218ca45c5d5ae5e86559b077cf1be213d4939eb1e9ab94eac4621d
-  data.tar.gz: 5862466e565d1e6030c35494b5028ae980a47d373e90050c62266055fcecd374
+  metadata.gz: 2bcd9baeafc1a271f75ccd74123f50ebd9d4fbe9065c2583f376c562f8e49155
+  data.tar.gz: 937dda6bbe4c41953f1e6eb1ea205eaa54277ae9f4202fa8a1e7e789348a76ad
 SHA512:
-  metadata.gz: 988d55c681a102e0c65b9133c6aeafc049e33755955f959d6e6046f5601dd192af881424355a2b373ed2e7a5a16b74236698aef5372e09584b10fe28d1b7bc21
-  data.tar.gz: adc58efa3b46d9fc1a87ddb2a4df32472507d61f21a3a0eb07026068cc5e41af166fb0a0f8ae23f1b23aec649b22835a50edbed79d35255e8cc231b82b31eb8c
+  metadata.gz: cbad4cc283bb449116b360bc4ef8002928add3399005bcc30aaccdf95ea03233f0d035862de643b4aa4d688eedbeaaa7dc029c67a2336156d7e03c9435468cfa
+  data.tar.gz: 83bfa0f53d7c0e094f271bfb3ddfef21ca58d41d77e1278886b5e26216a5b614629c9be33bc587bccc62e280612c75dbd0356fce772a727ed8cc003f86a03976

data/.coveralls.yml

@@ -0,0 +1 @@
+service_name: github-ci

data/.github/workflows/coverage.yml

@@ -0,0 +1,28 @@
+name: coverage
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+
+jobs:
+  coverage:
+    runs-on: ubuntu-20.04
+    steps:
+      - uses: actions/checkout@v2
+      - name: Install BLAS and LAPACK
+        run: sudo apt-get install -y libopenblas-dev liblapacke-dev
+      - name: Set up Ruby 2.7
+        uses: actions/setup-ruby@v1
+        with:
+          ruby-version: '2.7'
+      - name: Build and test with Rake
+        run: |
+          gem install bundler
+          bundle install
+          bundle exec rake
+      - name: Coveralls GitHub Action
+        uses: coverallsapp/github-action@v1.1.2
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}

data/.gitignore

@@ -16,6 +16,7 @@
 tags
 .DS_Store
 .ruby-version
+iterate.dat
 /spec/dump_dbl.t
 /spec/dump_int.t
 /spec/dump_mult_dbl.t

data/CHANGELOG.md

@@ -1,3 +1,12 @@
+# 0.22.3
+- Add regressor class for non-negative least square method.
+  - [NNLS](https://yoshoku.github.io/rumale/doc/Rumale/LinearModel/NNLS.html)
+- Add lbfgs solver to [Ridge](https://yoshoku.github.io/rumale/doc/Rumale/LinearModel/Ridge.html) and [LinearRegression](https://yoshoku.github.io/rumale/doc/Rumale/LinearModel/LinearRegression.html).
+  - In version 0.23.0, these classes will be changed to attempt to optimize with 'svd' or 'lbfgs' solver if 'auto' is given to
+  the solver parameter. If you use 'sgd' solver, you need specify it explicitly.
+- Add GC guard to native extension codes.
+- Update API documentation.
+
 # 0.22.2
 - Add classifier and regressor classes for stacking method.
   - [StackingClassifier](https://yoshoku.github.io/rumale/doc/Rumale/Ensemble/StackingClassifier.html)

data/Gemfile

@@ -13,4 +13,5 @@ gem 'rubocop', '~> 1.0'
 gem 'rubocop-performance', '~> 1.8'
 gem 'rubocop-rake', '~> 0.5'
 gem 'rubocop-rspec', '~> 2.0'
-gem 'simplecov', '~> 0.19'
+gem 'simplecov', '~> 0.21'
+gem 'simplecov-lcov', '~> 0.8'

data/LICENSE.txt

@@ -1,4 +1,4 @@
-Copyright (c) 2017-2020 Atsushi Tatsuma
+Copyright (c) 2017-2021 Atsushi Tatsuma
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without

data/README.md

@@ -3,6 +3,7 @@
 ![Rumale](https://dl.dropboxusercontent.com/s/joxruk2720ur66o/rumale_header_400.png)
 
 [![Build Status](https://github.com/yoshoku/rumale/workflows/build/badge.svg)](https://github.com/yoshoku/rumale/actions?query=workflow%3Abuild)
+[![Coverage Status](https://coveralls.io/repos/github/yoshoku/rumale/badge.svg?branch=main)](https://coveralls.io/github/yoshoku/rumale?branch=main)
 [![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/main/LICENSE.txt)
 [![Documentation](https://img.shields.io/badge/api-reference-blue.svg)](https://yoshoku.github.io/rumale/doc/)
@@ -176,7 +177,7 @@ For example, using the [OpenBLAS](https://github.com/xianyi/OpenBLAS) speeds up
 
 Install OpenBLAS library.
 
-Mac:
+macOS:
 
 ```bash
 $ brew install openblas
@@ -185,12 +186,13 @@ $ brew install openblas
 Ubuntu:
 
 ```bash
-$ sudo apt-get install gcc gfortran
-$ wget https://github.com/xianyi/OpenBLAS/archive/v0.3.5.tar.gz
-$ tar xzf v0.3.5.tar.gz
-$ cd OpenBLAS-0.3.5
-$ make USE_OPENMP=1
-$ sudo make PREFIX=/usr/local install
+$ sudo apt-get install libopenblas-dev liblapacke-dev
+```
+
+Windows (MSYS2):
+
+```bash
+$ pacman -S mingw-w64-x86_64-ruby mingw-w64-x86_64-openblas mingw-w64-x86_64-lapack
 ```
 
 Install Numo::Linalg gem.
@@ -206,6 +208,37 @@ require 'numo/linalg/autoloader'
 require 'rumale'
 ```
 
+### Numo::OpenBLAS
+[Numo::OpenBLAS](https://github.com/yoshoku/numo-openblas) downloads and builds OpenBLAS during installation
+and uses that as a background library for Numo::Linalg.
+
+Install compilers for building OpenBLAS.
+
+macOS:
+
+```bash
+$ brew install gcc gfortran make
+```
+
+Ubuntu:
+
+```bash
+$ sudo apt-get install gcc gfortran make
+```
+
+Install Numo::OpenBLAS gem.
+
+```bash
+$ gem install numo-openblas
+```
+
+Load Numo::OpenBLAS gem instead of Numo::Linalg.
+
+```ruby
+require 'numo/openblas'
+require 'rumale'
+```
+
 ### Parallel
 Several estimators in Rumale support parallel processing.
 Parallel processing in Rumale is realized by [Parallel](https://github.com/grosser/parallel) gem,
@@ -227,6 +260,10 @@ When -1 is given to n_jobs parameter, all processors are used.
 estimator = Rumale::Ensemble::RandomForestClassifier.new(n_jobs: -1, random_seed: 1)
 ```
 
+## Related Projects
+- [Rumale::SVM](https://github.com/yoshoku/rumale-svm) provides support vector machine algorithms in LIBSVM and LIBLINEAR with Rumale interface.
+- [Rumale::Torch](https://github.com/yoshoku/rumale-torch) provides the learning and inference by the neural network defined in torch.rb with Rumale interface.
+
 ## Novelties
 
 * [Rumale SHOP](https://suzuri.jp/yoshoku)

data/ext/rumale/tree.c

@@ -257,10 +257,13 @@ find_split_params_cls(VALUE self, VALUE criterion, VALUE impurity, VALUE order,
   split_opts_cls opts = { StringValuePtr(criterion), NUM2LONG(n_classes), NUM2DBL(impurity) };
   VALUE params = na_ndloop3(&ndf, &opts, 3, order, features, labels);
   VALUE results = rb_ary_new2(4);
-  rb_ary_store(results, 0, DBL2NUM(((double*)na_get_pointer_for_read(params))[0]));
-  rb_ary_store(results, 1, DBL2NUM(((double*)na_get_pointer_for_read(params))[1]));
-  rb_ary_store(results, 2, DBL2NUM(((double*)na_get_pointer_for_read(params))[2]));
-  rb_ary_store(results, 3, DBL2NUM(((double*)na_get_pointer_for_read(params))[3]));
+  double* params_ptr = (double*)na_get_pointer_for_read(params);
+  rb_ary_store(results, 0, DBL2NUM(params_ptr[0]));
+  rb_ary_store(results, 1, DBL2NUM(params_ptr[1]));
+  rb_ary_store(results, 2, DBL2NUM(params_ptr[2]));
+  rb_ary_store(results, 3, DBL2NUM(params_ptr[3]));
+  RB_GC_GUARD(params);
+  RB_GC_GUARD(criterion);
   return results;
 }
 
@@ -375,10 +378,13 @@ find_split_params_reg(VALUE self, VALUE criterion, VALUE impurity, VALUE order,
   split_opts_reg opts = { StringValuePtr(criterion), NUM2DBL(impurity) };
   VALUE params = na_ndloop3(&ndf, &opts, 3, order, features, targets);
   VALUE results = rb_ary_new2(4);
-  rb_ary_store(results, 0, DBL2NUM(((double*)na_get_pointer_for_read(params))[0]));
-  rb_ary_store(results, 1, DBL2NUM(((double*)na_get_pointer_for_read(params))[1]));
-  rb_ary_store(results, 2, DBL2NUM(((double*)na_get_pointer_for_read(params))[2]));
-  rb_ary_store(results, 3, DBL2NUM(((double*)na_get_pointer_for_read(params))[3]));
+  double* params_ptr = (double*)na_get_pointer_for_read(params);
+  rb_ary_store(results, 0, DBL2NUM(params_ptr[0]));
+  rb_ary_store(results, 1, DBL2NUM(params_ptr[1]));
+  rb_ary_store(results, 2, DBL2NUM(params_ptr[2]));
+  rb_ary_store(results, 3, DBL2NUM(params_ptr[3]));
+  RB_GC_GUARD(params);
+  RB_GC_GUARD(criterion);
   return results;
 }
 
@@ -464,8 +470,10 @@ find_split_params_grad_reg
   double opts[3] = { NUM2DBL(sum_gradient), NUM2DBL(sum_hessian), NUM2DBL(reg_lambda) };
   VALUE params = na_ndloop3(&ndf, opts, 4, order, features, gradients, hessians);
   VALUE results = rb_ary_new2(2);
-  rb_ary_store(results, 0, DBL2NUM(((double*)na_get_pointer_for_read(params))[0]));
-  rb_ary_store(results, 1, DBL2NUM(((double*)na_get_pointer_for_read(params))[1]));
+  double* params_ptr = (double*)na_get_pointer_for_read(params);
+  rb_ary_store(results, 0, DBL2NUM(params_ptr[0]));
+  rb_ary_store(results, 1, DBL2NUM(params_ptr[1]));
+  RB_GC_GUARD(params);
   return results;
 }
 
@@ -497,6 +505,9 @@ node_impurity_cls(VALUE self, VALUE criterion, VALUE y_nary, VALUE n_elements_,
 
   xfree(histogram);
 
+  RB_GC_GUARD(y_nary);
+  RB_GC_GUARD(criterion);
+
   return ret;
 }
 
@@ -531,6 +542,8 @@ node_impurity_reg(VALUE self, VALUE criterion, VALUE y)
 
   xfree(sum_vec);
 
+  RB_GC_GUARD(criterion);
+
   return ret;
 }
 

data/lib/rumale.rb

@@ -30,6 +30,7 @@ require 'rumale/linear_model/linear_regression'
 require 'rumale/linear_model/ridge'
 require 'rumale/linear_model/lasso'
 require 'rumale/linear_model/elastic_net'
+require 'rumale/linear_model/nnls'
 require 'rumale/kernel_machine/kernel_svc'
 require 'rumale/kernel_machine/kernel_pca'
 require 'rumale/kernel_machine/kernel_fda'

data/lib/rumale/base/base_estimator.rb

@@ -11,13 +11,15 @@ module Rumale
 
       private
 
-      def enable_linalg?
+      def enable_linalg?(warning: true)
         if defined?(Numo::Linalg).nil?
-          warn('If you want to use features that depend on Numo::Linalg, you should install and load Numo::Linalg in advance.')
+          warn('If you want to use features that depend on Numo::Linalg, you should install and load Numo::Linalg in advance.') if warning
           return false
         end
         if Numo::Linalg::VERSION < '0.1.4'
-          warn('The loaded Numo::Linalg does not implement the methods required by Rumale. Please load Numo::Linalg version 0.1.4 or later.')
+          if warning
+            warn('The loaded Numo::Linalg does not implement the methods required by Rumale. Please load Numo::Linalg version 0.1.4 or later.')
+          end
           return false
         end
         true

data/lib/rumale/linear_model/elastic_net.rb

@@ -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

@@ -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

@@ -6,7 +6,8 @@ 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 =
@@ -41,31 +42,32 @@ 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: 1000, batch_size: 50, tol: 1e-4,
@@ -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' # rubocop:disable Style/NegatedIfElseCondition
+                             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/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

@@ -1,12 +1,15 @@
 # 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 =
@@ -41,32 +44,33 @@ 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,
@@ -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' # rubocop:disable Style/NegatedIfElseCondition
+                             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/validation.rb

@@ -27,6 +27,7 @@ module Rumale
       y
     end
 
+    # @deprecated Use check_convert_sample_array instead of this method.
     # @!visibility private
     def check_sample_array(x)
       raise TypeError, 'Expect class of sample matrix to be Numo::DFloat' unless x.is_a?(Numo::DFloat)
@@ -35,6 +36,7 @@ module Rumale
       nil
     end
 
+    # @deprecated Use check_convert_label_array instead of this method.
     # @!visibility private
     def check_label_array(y)
       raise TypeError, 'Expect class of label vector to be Numo::Int32' unless y.is_a?(Numo::Int32)
@@ -43,6 +45,7 @@ module Rumale
       nil
     end
 
+    # @deprecated Use check_convert_tvalue_array instead of this method.
     # @!visibility private
     def check_tvalue_array(y)
       raise TypeError, 'Expect class of target value vector to be Numo::DFloat' unless y.is_a?(Numo::DFloat)
@@ -64,49 +67,58 @@ module Rumale
       nil
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_type(type, params = {})
       params.each { |k, v| raise TypeError, "Expect class of #{k} to be #{type}" unless v.is_a?(type) }
       nil
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_type_or_nil(type, params = {})
       params.each { |k, v| raise TypeError, "Expect class of #{k} to be #{type} or nil" unless v.is_a?(type) || v.is_a?(NilClass) }
       nil
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_numeric(params = {})
       check_params_type(Numeric, params)
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_numeric_or_nil(params = {})
       check_params_type_or_nil(Numeric, params)
     end
 
+    # @deprecated Use check_params_numeric instead of this method.
     # @!visibility private
     def check_params_float(params = {})
       check_params_type(Float, params)
     end
 
+    # @deprecated Use check_params_numeric instead of this method.
     # @!visibility private
     def check_params_integer(params = {})
       check_params_type(Integer, params)
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_string(params = {})
       check_params_type(String, params)
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_boolean(params = {})
       params.each { |k, v| raise TypeError, "Expect class of #{k} to be Boolean" unless v.is_a?(FalseClass) || v.is_a?(TrueClass) }
       nil
     end
 
+    # TODO: Better to replace with RBS in the future.
     # @!visibility private
     def check_params_positive(params = {})
       params.compact.each { |k, v| raise ArgumentError, "Expect #{k} to be positive value" if v.negative? }

data/lib/rumale/version.rb

@@ -3,5 +3,5 @@
 # Rumale is a machine learning library in Ruby.
 module Rumale
   # The version of Rumale you are using.
-  VERSION = '0.22.2'
+  VERSION = '0.22.3'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: rumale
 version: !ruby/object:Gem::Version
-  version: 0.22.2
+  version: 0.22.3
 platform: ruby
 authors:
 - yoshoku
-autorequire: 
+autorequire:
 bindir: exe
 cert_chain: []
-date: 2021-01-10 00:00:00.000000000 Z
+date: 2021-01-23 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: numo-narray
@@ -57,7 +57,9 @@ extensions:
 - ext/rumale/extconf.rb
 extra_rdoc_files: []
 files:
+- ".coveralls.yml"
 - ".github/workflows/build.yml"
+- ".github/workflows/coverage.yml"
 - ".gitignore"
 - ".rspec"
 - ".rubocop.yml"
@@ -141,6 +143,7 @@ files:
 - lib/rumale/linear_model/lasso.rb
 - lib/rumale/linear_model/linear_regression.rb
 - lib/rumale/linear_model/logistic_regression.rb
+- lib/rumale/linear_model/nnls.rb
 - lib/rumale/linear_model/ridge.rb
 - lib/rumale/linear_model/svc.rb
 - lib/rumale/linear_model/svr.rb
@@ -211,7 +214,7 @@ metadata:
   source_code_uri: https://github.com/yoshoku/rumale
   documentation_uri: https://yoshoku.github.io/rumale/doc/
   bug_tracker_uri: https://github.com/yoshoku/rumale/issues
-post_install_message: 
+post_install_message:
 rdoc_options: []
 require_paths:
 - lib
@@ -226,8 +229,8 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.1.4
-signing_key: 
+rubygems_version: 3.2.3
+signing_key:
 specification_version: 4
 summary: Rumale is a machine learning library in Ruby. Rumale provides machine learning
   algorithms with interfaces similar to Scikit-Learn in Python.