rumale 0.21.0 → 0.22.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: fd5ca16629a5258be9e577771dc8c6b42dbfdf3a60a4c43d4ee170cc17b72bea
-  data.tar.gz: 70b26bcf0b39bb8e716b9bbb4aba100c496152b1c4e71879105046440c7d8758
+  metadata.gz: 4e2f68b3182ada73537901e7bc74bddd100aff75264f9147c88d8240fb624e29
+  data.tar.gz: e2639a55fc84d1399b925f65b3a56b38f2ae3150dd15ab8556120af28d408cae
 SHA512:
-  metadata.gz: 645a6bda6e3601534c69f5ecfbd840c1d6c1ed7a5a3b8bd57995621a03d970cd02e9749a8a70be5af2678c029a26a5e6e1c32376a4514a64e96d6a9b4b12aa3e
-  data.tar.gz: 5904c64da9cc30cf0c288dfbeb3051bca3333e588e153cf19619c169d713e93edcb95e6902134e58c823d621ebad9e6a56310123c110b6d810033f5f96a40fbb
+  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/CHANGELOG.md

@@ -1,3 +1,9 @@
+# 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.

data/Gemfile

@@ -3,9 +3,7 @@ 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'
@@ -14,3 +12,4 @@ 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/)

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: 1000, 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: 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

data/lib/rumale/metric_learning/neighbourhood_component_analysis.rb

@@ -2,13 +2,13 @@
 
 require 'rumale/base/base_estimator'
 require 'rumale/base/transformer'
+require 'lbfgsb'
 
 module Rumale
   module MetricLearning
     # NeighbourhoodComponentAnalysis is a class that implements Neighbourhood Component Analysis.
     #
     # @example
-    #   require 'mopti'
     #   require 'rumale'
     #
     #   transformer = Rumale::MetricLearning::NeighbourhoodComponentAnalysis.new
@@ -39,7 +39,9 @@ module Rumale
       # @param init [String] The initialization method for components ('random' or 'pca').
       # @param max_iter [Integer] The maximum number of iterations.
       # @param tol [Float] The tolerance of termination criterion.
+      #   This value is given as tol / Lbfgsb::DBL_EPSILON to the factr argument of Lbfgsb.minimize method.
       # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
+      #   If true is given, 'iterate.dat' file is generated by lbfgsb.rb.
       # @param random_seed [Integer] The seed value using to initialize the random generator.
       def initialize(n_components: nil, init: 'random', max_iter: 100, tol: 1e-6, verbose: false, random_seed: nil)
         check_params_numeric_or_nil(n_components: n_components, random_seed: random_seed)
@@ -65,8 +67,6 @@ module Rumale
       # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
       # @return [NeighbourhoodComponentAnalysis] The learned classifier itself.
       def fit(x, y)
-        raise 'NeighbourhoodComponentAnalysis#fit requires Mopti but that is not loaded.' unless enable_mopti?
-
         x = check_convert_sample_array(x)
         y = check_convert_label_array(y)
         check_sample_label_size(x, y)
@@ -102,14 +102,6 @@ module Rumale
 
       private
 
-      def enable_mopti?
-        if defined?(Mopti).nil?
-          warn('NeighbourhoodComponentAnalysis#fit requires Mopti but that is not loaded. You should intall and load mopti gem in advance.')
-          return false
-        end
-        true
-      end
-
       def init_components(x, n_features, n_components)
         if @params[:init] == 'pca'
           pca = Rumale::Decomposition::PCA.new(n_components: n_components)
@@ -127,28 +119,18 @@ module Rumale
         res[:x] = comp_init
         res[:n_iter] = 0
         # perform optimization.
-        optimizer = Mopti::ScaledConjugateGradient.new(
-          fnc: method(:nca_loss), jcb: method(:nca_dloss),
-          x_init: comp_init, args: [x, y],
-          max_iter: @params[:max_iter], ftol: @params[:tol]
+        verbose = @params[:verbose] ? 1 : -1
+        res = Lbfgsb.minimize(
+          fnc: method(:nca_fnc), jcb: true, x_init: comp_init, args: [x, y],
+          maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON, verbose: verbose
         )
-        fold = 0.0
-        dold = 0.0
-        optimizer.each do |prm|
-          res = prm
-          puts "[NeighbourhoodComponentAnalysis] The value of objective function after #{res[:n_iter]} epochs: #{x.shape[0] - res[:fnc]}" if @params[:verbose]
-          break if (fold - res[:fnc]).abs <= @params[:tol] && (dold - res[:jcb]).abs <= @params[:tol]
-
-          fold = res[:fnc]
-          dold = res[:jcb]
-        end
         # return the results.
         n_iter = res[:n_iter]
         comps = n_components == 1 ? res[:x].dup : res[:x].reshape(n_components, n_features)
         [comps, n_iter]
       end
 
-      def nca_loss(w, x, y)
+      def nca_fnc(w, x, y)
         # initialize some variables.
         n_samples, n_features = x.shape
         n_components = w.size / n_features
@@ -157,32 +139,18 @@ module Rumale
         z = x.dot(w.transpose)
         # calculate probability matrix.
         prob_mat = probability_matrix(z)
-        # calculate loss.
+        # calculate loss and gradient.
         # NOTE:
         # NCA attempts to maximize its objective function.
         # For the minization algorithm, the objective function value is subtracted from the maixmum value (n_samples).
         mask_mat = y.expand_dims(1).eq(y)
         masked_prob_mat = prob_mat * mask_mat
-        n_samples - masked_prob_mat.sum
-      end
-
-      def nca_dloss(w, x, y)
-        # initialize some variables.
-        n_features = x.shape[1]
-        n_components = w.size / n_features
-        # projection.
-        w = w.reshape(n_components, n_features)
-        z = x.dot(w.transpose)
-        # calculate probability matrix.
-        prob_mat = probability_matrix(z)
-        # calculate gradient.
-        mask_mat = y.expand_dims(1).eq(y)
-        masked_prob_mat = prob_mat * mask_mat
+        loss = n_samples - masked_prob_mat.sum
         weighted_prob_mat = masked_prob_mat - prob_mat * masked_prob_mat.sum(1).expand_dims(1)
         weighted_prob_mat += weighted_prob_mat.transpose
         weighted_prob_mat[weighted_prob_mat.diag_indices] = -weighted_prob_mat.sum(0)
-        gradient = 2 * z.transpose.dot(weighted_prob_mat).dot(x)
-        -gradient.flatten.dup
+        gradient = -2 * z.transpose.dot(weighted_prob_mat).dot(x)
+        [loss, gradient.flatten.dup]
       end
 
       def probability_matrix(z)

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 = '0.21.0'
+  VERSION = '0.22.0'
 end

data/rumale.gemspec

@@ -45,4 +45,5 @@ Gem::Specification.new do |spec|
   }
 
   spec.add_runtime_dependency 'numo-narray', '>= 0.9.1'
+  spec.add_runtime_dependency 'lbfgsb', '>=0.3.0'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: rumale
 version: !ruby/object:Gem::Version
-  version: 0.21.0
+  version: 0.22.0
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2020-10-02 00:00:00.000000000 Z
+date: 2020-11-22 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: numo-narray
@@ -24,6 +24,20 @@ dependencies:
     - - ">="
       - !ruby/object:Gem::Version
         version: 0.9.1
+- !ruby/object:Gem::Dependency
+  name: lbfgsb
+  requirement: !ruby/object:Gem::Requirement
+    requirements:
+    - - ">="
+      - !ruby/object:Gem::Version
+        version: 0.3.0
+  type: :runtime
+  prerelease: false
+  version_requirements: !ruby/object:Gem::Requirement
+    requirements:
+    - - ">="
+      - !ruby/object:Gem::Version
+        version: 0.3.0
 description: |
   Rumale is a machine learning library in Ruby.
   Rumale provides machine learning algorithms with interfaces similar to Scikit-Learn in Python.
@@ -43,7 +57,7 @@ extensions:
 - ext/rumale/extconf.rb
 extra_rdoc_files: []
 files:
-- ".coveralls.yml"
+- ".github/workflows/build.yml"
 - ".gitignore"
 - ".rspec"
 - ".rubocop.yml"
@@ -209,7 +223,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.1.2
+rubygems_version: 3.1.4
 signing_key:
 specification_version: 4
 summary: Rumale is a machine learning library in Ruby. Rumale provides machine learning

data/.coveralls.yml

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