svc.rb in rumale-linear_model-0.25.0

- old
+ new

@@ -1,17 +1,18 @@
 # frozen_string_literal: true
 
+require 'lbfgsb'
+
 require 'rumale/base/classifier'
-require 'rumale/linear_model/base_sgd'
 require 'rumale/probabilistic_output'
 require 'rumale/validation'
 
+require_relative 'base_estimator'
+
 module Rumale
-  # This module consists of the classes that implement generalized linear models.
   module LinearModel
-    # SVC is a class that implements Support Vector Classifier
-    # with stochastic gradient descent optimization.
+    # SVC is a class that implements Support Vector Classifier with the squared hinge loss.
     # For multiclass classification problem, it uses one-vs-the-rest strategy.
     #
     # @note
     #   Rumale::SVM provides linear support vector classifier based on LIBLINEAR.
     #   If you prefer execution speed, you should use Rumale::SVM::LinearSVC.
@@ -19,99 +20,70 @@
     #
     # @example
     #   require 'rumale/linear_model/svc'
     #
     #   estimator =
-    #     Rumale::LinearModel::SVC.new(reg_param: 1.0, max_iter: 1000, batch_size: 50, random_seed: 1)
+    #     Rumale::LinearModel::SVC.new(reg_param: 1.0)
     #   estimator.fit(training_samples, traininig_labels)
     #   results = estimator.predict(testing_samples)
-    #
-    # *Reference*
-    # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
-    # - Tsuruoka, Y., Tsujii, J., and Ananiadou, S., "Stochastic Gradient Descent Training for L1-regularized Log-linear Models with Cumulative Penalty," Proc. ACL'09, pp. 477--485, 2009.
-    # - Bottou, L., "Large-Scale Machine Learning with Stochastic Gradient Descent," Proc. COMPSTAT'10, pp. 177--186, 2010.
-    class SVC < BaseSGD
-      include ::Rumale::Base::Classifier
+    class SVC < Rumale::LinearModel::BaseEstimator
+      include Rumale::Base::Classifier
 
-      # Return the weight vector for SVC.
-      # @return [Numo::DFloat] (shape: [n_classes, n_features])
-      attr_reader :weight_vec
-
-      # Return the bias term (a.k.a. intercept) for SVC.
-      # @return [Numo::DFloat] (shape: [n_classes])
-      attr_reader :bias_term
-
       # Return the class labels.
       # @return [Numo::Int32] (shape: [n_classes])
       attr_reader :classes
 
-      # Return the random generator for performing random sampling.
-      # @return [Random]
-      attr_reader :rng
-
-      # Create a new classifier with Support Vector Machine by the SGD optimization.
+      # Create a new linear classifier with Support Vector Machine with the squared hinge loss.
       #
-      # @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).
-      # @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'.
-      # @param momentum [Float] The momentum factor.
-      # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
-      # @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.
       # @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.
-      # @param batch_size [Integer] The size of the mini batches.
       # @param tol [Float] The tolerance of loss for terminating optimization.
       # @param probability [Boolean] The flag indicating whether to perform probability estimation.
       # @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.
       # @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(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,
-                     probability: false,
-                     n_jobs: nil, verbose: false, random_seed: nil)
+      #   'iterate.dat' file is generated by lbfgsb.rb.
+      def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0, max_iter: 1000, tol: 1e-4, probability: false,
+                     n_jobs: nil, verbose: false)
         super()
-        @params.merge!(method(:initialize).parameters.to_h { |_t, arg| [arg, binding.local_variable_get(arg)] })
-        @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
-        @params[:random_seed] ||= srand
-        @rng = Random.new(@params[:random_seed])
-        @penalty_type = @params[:penalty]
-        @loss_func = ::Rumale::LinearModel::Loss::HingeLoss.new
+        @params = {
+          reg_param: reg_param,
+          fit_bias: fit_bias,
+          bias_scale: bias_scale,
+          max_iter: max_iter,
+          tol: tol,
+          probability: probability,
+          n_jobs: n_jobs,
+          verbose: verbose
+        }
       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::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
       # @return [SVC] The learned classifier itself.
       def fit(x, y)
-        x = ::Rumale::Validation.check_convert_sample_array(x)
-        y = ::Rumale::Validation.check_convert_label_array(y)
-        ::Rumale::Validation.check_sample_size(x, y)
+        x = Rumale::Validation.check_convert_sample_array(x)
+        y = Rumale::Validation.check_convert_label_array(y)
+        Rumale::Validation.check_sample_size(x, y)
 
         @classes = Numo::Int32[*y.to_a.uniq.sort]
+        x = expand_feature(x) if fit_bias?
 
         if multiclass_problem?
           n_classes = @classes.size
           n_features = x.shape[1]
-          # initialize model.
+          n_features -= 1 if fit_bias?
           @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
           @bias_term = Numo::DFloat.zeros(n_classes)
           @prob_param = Numo::DFloat.zeros(n_classes, 2)
-          # fit model.
           models = if enable_parallel?
                      parallel_map(n_classes) do |n|
                        bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
                        partial_fit(x, bin_y)
                      end
@@ -119,11 +91,10 @@
                      Array.new(n_classes) do |n|
                        bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
                        partial_fit(x, bin_y)
                      end
                    end
-          # store model.
           models.each_with_index { |model, n| @weight_vec[n, true], @bias_term[n], @prob_param[n, true] = model }
         else
           negative_label = @classes[0]
           bin_y = Numo::Int32.cast(y.ne(negative_label)) * 2 - 1
           @weight_vec, @bias_term, @prob_param = partial_fit(x, bin_y)
@@ -135,21 +106,21 @@
       # Calculate confidence scores for samples.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
       # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Confidence score per sample.
       def decision_function(x)
-        x = ::Rumale::Validation.check_convert_sample_array(x)
+        x = Rumale::Validation.check_convert_sample_array(x)
 
         x.dot(@weight_vec.transpose) + @bias_term
       end
 
       # Predict class labels for samples.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
       # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
       def predict(x)
-        x = ::Rumale::Validation.check_convert_sample_array(x)
+        x = Rumale::Validation.check_convert_sample_array(x)
 
         n_samples = x.shape[0]
         predicted = if multiclass_problem?
                       decision_values = decision_function(x)
                       if enable_parallel?
@@ -167,33 +138,54 @@
       # Predict probability for samples.
       #
       # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
       # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
       def predict_proba(x)
-        x = ::Rumale::Validation.check_convert_sample_array(x)
+        x = Rumale::Validation.check_convert_sample_array(x)
 
         if multiclass_problem?
           probs = 1.0 / (Numo::NMath.exp(@prob_param[true, 0] * decision_function(x) + @prob_param[true, 1]) + 1.0)
           (probs.transpose / probs.sum(axis: 1)).transpose.dup
         else
-          n_samples, = x.shape
+          n_samples = x.shape[0]
           probs = Numo::DFloat.zeros(n_samples, 2)
           probs[true, 1] = 1.0 / (Numo::NMath.exp(@prob_param[0] * decision_function(x) + @prob_param[1]) + 1.0)
           probs[true, 0] = 1.0 - probs[true, 1]
           probs
         end
       end
 
       private
 
-      def partial_fit(x, bin_y)
-        w, b = super
-        p = if @params[:probability]
-              ::Rumale::ProbabilisticOutput.fit_sigmoid(x.dot(w.transpose) + b, bin_y)
-            else
-              Numo::DFloat[1, 0]
-            end
-        [w, b, p]
+      def partial_fit(base_x, bin_y)
+        fnc = proc do |w, x, y, reg_param|
+          n_samples = x.shape[0]
+          z = x.dot(w)
+          t = 1 - y * z
+          loss = 0.5 * reg_param * w.dot(w) + (x.class.maximum(0, t)**2).sum.fdiv(n_samples)
+          indices = t.gt(0)
+          grad = reg_param * w
+          if indices.count.positive?
+            sx = x[indices, true]
+            sy = y[indices]
+            grad += 2.fdiv(n_samples) * sx.transpose.dot((sx.dot(w) - sy))
+          end
+          [loss, grad]
+        end
+
+        n_features = base_x.shape[1]
+        w_init = Numo::DFloat.zeros(n_features)
+
+        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: @params[:verbose] ? 1 : -1
+        )
+
+        prb = @params[:probability] ? Rumale::ProbabilisticOutput.fit_sigmoid(base_x.dot(res[:x]), bin_y) : Numo::DFloat[1, 0]
+        w, b = split_weight(res[:x])
+
+        [w, b, prb]
       end
 
       def multiclass_problem?
         @classes.size > 2
       end