rumale-svm 0.1.0

data/lib/rumale/svm/linear_svr.rb

@@ -0,0 +1,150 @@
+# frozen_string_literal: true
+
+require 'numo/libsvm'
+require 'rumale/base/base_estimator'
+require 'rumale/base/regressor'
+
+module Rumale
+  module SVM
+    # LinearSVR is a class that provides Support Vector Regressor in LIBLINEAR with Rumale interface.
+    #
+    # @example
+    #   estimator = Rumale::SVM::LinearSVR.new(reg_param: 1.0, random_seed: 1)
+    #   estimator.fit(training_samples, traininig_target_values)
+    #   results = estimator.predict(testing_samples)
+    class LinearSVR
+      include Base::BaseEstimator
+      include Base::Regressor
+
+      # Return the weight vector for LinearSVR.
+      # @return [Numo::DFloat] (shape: [n_classes, n_features])
+      attr_reader :weight_vec
+
+      # Return the bias term (a.k.a. intercept) for LinearSVR.
+      # @return [Numo::DFloat] (shape: [n_classes])
+      attr_reader :bias_term
+
+      # Create a new regressor with Support Vector Regressor.
+      #
+      # @param loss [String] The type of loss function ('squared_epsilon_insensitive' or 'epsilon_insensitive').
+      # @param dual [Boolean] The flag indicating whether to solve dual optimization problem.
+      #   When n_samples > n_features, dual = false is more preferable.
+      #   This parameter is ignored if loss = 'epsilon_insensitive'.
+      # @param reg_param [Float] The regularization parameter.
+      # @param epsilon [Float] The epsilon parameter in loss function of espsilon-svr.
+      # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
+      # @param bias_scale [Float] The scale of the bias term.
+      #   This parameter is ignored if fit_bias = false.
+      # @param tol [Float] The tolerance of termination criterion.
+      # @param verbose [Boolean] The flag indicating whether to output learning process message
+      # @param random_seed [Integer/Nil] The seed value using to initialize the random generator.
+      def initialize(loss: 'squared_epsilon_insensitive', dual: true, reg_param: 1.0, epsilon: 0.1,
+                     fit_bias: true, bias_scale: 1.0, tol: 1e-3, verbose: false, random_seed: nil)
+        check_params_string(loss: loss)
+        check_params_float(reg_param: reg_param, epsilon: epsilon, bias_scale: bias_scale, tol: tol)
+        check_params_boolean(dual: dual, fit_bias: fit_bias, verbose: verbose)
+        check_params_type_or_nil(Integer, random_seed: random_seed)
+        @params = {}
+        @params[:loss] = loss == 'epsilon_insensitive' ? 'epsilon_insensitive' : 'squared_epsilon_insensitive'
+        @params[:dual] = dual
+        @params[:reg_param] = reg_param
+        @params[:epsilon] = epsilon
+        @params[:fit_bias] = fit_bias
+        @params[:bias_scale] = bias_scale
+        @params[:tol] = tol
+        @params[:verbose] = verbose
+        @params[:random_seed] = random_seed
+        @model = nil
+        @weight_vec = nil
+        @bias_term = nil
+      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]) The target values to be used for fitting the model.
+      # @return [LinearSVR] The learned regressor itself.
+      def fit(x, y)
+        check_sample_array(x)
+        check_tvalue_array(y)
+        check_sample_tvalue_size(x, y)
+        xx = fit_bias? ? expand_feature(x) : x
+        @model = Numo::Liblinear.train(xx, y, liblinear_params)
+        @weight_vec, @bias_term = weight_and_bias(@model[:w])
+        self
+      end
+
+      # Predict values for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
+      # @return [Numo::DFloat] (shape: [n_samples]) Predicted value per sample.
+      def predict(x)
+        check_sample_array(x)
+        xx = fit_bias? ? expand_feature(x) : x
+        Numo::Liblinear.predict(xx, liblinear_params, @model)
+      end
+
+      # Dump marshal data.
+      # @return [Hash] The marshal data about LinearSVR.
+      def marshal_dump
+        { params: @params,
+          model: @model,
+          weight_vec: @weight_vec,
+          bias_term: @bias_term }
+      end
+
+      # Load marshal data.
+      # @return [nil]
+      def marshal_load(obj)
+        @params = obj[:params]
+        @model = obj[:model]
+        @weight_vec = obj[:weight_vec]
+        @bias_term = obj[:bias_term]
+        nil
+      end
+
+      private
+
+      def expand_feature(x)
+        n_samples = x.shape[0]
+        Numo::NArray.hstack([x, Numo::DFloat.ones([n_samples, 1]) * bias_scale])
+      end
+
+      def weight_and_bias(base_weight)
+        bias_vec = 0.0
+        weight_mat = base_weight.dup
+        if fit_bias?
+          bias_vec = weight_mat[-1]
+          weight_mat = weight_mat[0...-1].dup
+        end
+        [weight_mat, bias_vec]
+      end
+
+      def liblinear_params
+        res = {}
+        res[:solver_type] = solver_type
+        res[:eps] = @params[:tol]
+        res[:C] = @params[:reg_param]
+        res[:p] = @params[:epsilon]
+        res[:verbose] = @params[:verbose]
+        res[:random_seed] = @params[:random_seed]
+        res
+      end
+
+      def solver_type
+        return Numo::Liblinear::SolverType::L2R_L1LOSS_SVR_DUAL if @params[:loss] == 'epsilon_insensitive'
+        return Numo::Liblinear::SolverType::L2R_L2LOSS_SVR_DUAL if @params[:dual]
+
+        Numo::Liblinear::SolverType::L2R_L2LOSS_SVR
+      end
+
+      def fit_bias?
+        @params[:fit_bias]
+      end
+
+      def bias_scale
+        @params[:bias_scale]
+      end
+    end
+  end
+end

data/lib/rumale/svm/logistic_regression.rb

@@ -0,0 +1,190 @@
+# frozen_string_literal: true
+
+require 'numo/liblinear'
+require 'rumale/base/base_estimator'
+require 'rumale/base/classifier'
+
+module Rumale
+  module SVM
+    # LogisticRegression is a class that provides Logistic Regression in LIBLINEAR with Rumale interface
+    #
+    # @example
+    #   estimator = Rumale::SVM::LogisticRegression.new(penalty: 'l2', dual: false, reg_param: 1.0, random_seed: 1)
+    #   estimator.fit(training_samples, traininig_labels)
+    #   results = estimator.predict(testing_samples)
+    class LogisticRegression
+      include Base::BaseEstimator
+      include Base::Classifier
+
+      # Return the weight vector for LogisticRegression.
+      # @return [Numo::DFloat] (shape: [n_classes, n_features])
+      attr_reader :weight_vec
+
+      # Return the bias term (a.k.a. intercept) for LogisticRegression.
+      # @return [Numo::DFloat] (shape: [n_classes])
+      attr_reader :bias_term
+
+      # Create a new classifier with Logistic Regression.
+      #
+      # @param penalty [String] The type of norm used in the penalization ('l2' or 'l1').
+      # @param dual [Boolean] The flag indicating whether to solve dual optimization problem.
+      #   When n_samples > n_features, dual = false is more preferable.
+      #   This parameter is ignored if penalty = 'l1'.
+      # @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.
+      #   This parameter is ignored if fit_bias = false.
+      # @param tol [Float] The tolerance of termination criterion.
+      # @param verbose [Boolean] The flag indicating whether to output learning process message
+      # @param random_seed [Integer/Nil] The seed value using to initialize the random generator.
+      def initialize(penalty: 'l2', dual: true, reg_param: 1.0,
+                     fit_bias: true, bias_scale: 1.0,
+                     tol: 1e-3, verbose: false, random_seed: nil)
+        check_params_string(penalty: penalty)
+        check_params_float(reg_param: reg_param, bias_scale: bias_scale, tol: tol)
+        check_params_boolean(dual: dual, fit_bias: fit_bias, verbose: verbose)
+        check_params_type_or_nil(Integer, random_seed: random_seed)
+        @params = {}
+        @params[:penalty] = penalty == 'l1' ? 'l1' : 'l2'
+        @params[:dual] = dual
+        @params[:reg_param] = reg_param
+        @params[:fit_bias] = fit_bias
+        @params[:bias_scale] = bias_scale
+        @params[:tol] = tol
+        @params[:verbose] = verbose
+        @params[:random_seed] = random_seed
+        @model = nil
+        @weight_vec = nil
+        @bias_term = nil
+      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 [LogisticRegression] The learned classifier itself.
+      def fit(x, y)
+        check_sample_array(x)
+        check_label_array(y)
+        check_sample_label_size(x, y)
+        xx = fit_bias? ? expand_feature(x) : x
+        @model = Numo::Liblinear.train(xx, y, liblinear_params)
+        @weight_vec, @bias_term = weight_and_bias(@model[:w])
+        self
+      end
+
+      # 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)
+        check_sample_array(x)
+        xx = fit_bias? ? expand_feature(x) : x
+        Numo::Liblinear.decision_function(xx, liblinear_params, @model)
+      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)
+        check_sample_array(x)
+        xx = fit_bias? ? expand_feature(x) : x
+        Numo::Int32.cast(Numo::Liblinear.predict(xx, liblinear_params, @model))
+      end
+
+      # Predict class probability for samples.
+      # This method works correctly only if the probability parameter is true.
+      #
+      # @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)
+        check_sample_array(x)
+        xx = fit_bias? ? expand_feature(x) : x
+        Numo::Liblinear.predict_proba(xx, liblinear_params, @model)
+      end
+
+      # Dump marshal data.
+      # @return [Hash] The marshal data about LogisticRegression.
+      def marshal_dump
+        { params: @params,
+          model: @model,
+          weight_vec: @weight_vec,
+          bias_term: @bias_term }
+      end
+
+      # Load marshal data.
+      # @return [nil]
+      def marshal_load(obj)
+        @params = obj[:params]
+        @model = obj[:model]
+        @weight_vec = obj[:weight_vec]
+        @bias_term = obj[:bias_term]
+        nil
+      end
+
+      private
+
+      def expand_feature(x)
+        n_samples = x.shape[0]
+        Numo::NArray.hstack([x, Numo::DFloat.ones([n_samples, 1]) * bias_scale])
+      end
+
+      def weight_and_bias(base_weight)
+        if binary_class?
+          bias_vec = 0.0
+          weight_mat = base_weight.dup
+          if fit_bias?
+            bias_vec = weight_mat[-1]
+            weight_mat = weight_mat[0...-1].dup
+          end
+        else
+          bias_vec = Numo::DFloat.zeros(n_classes)
+          weight_mat = base_weight.reshape(n_features, n_classes).transpose.dup
+          if fit_bias?
+            bias_vec = weight_mat[true, -1].dup
+            weight_mat = weight_mat[true, 0...-1].dup
+          end
+        end
+        [weight_mat, bias_vec]
+      end
+
+      def liblinear_params
+        res = {}
+        res[:solver_type] = solver_type
+        res[:eps] = @params[:tol]
+        res[:C] = @params[:reg_param]
+        res[:verbose] = @params[:verbose]
+        res[:random_seed] = @params[:random_seed]
+        res
+      end
+
+      def solver_type
+        return Numo::Liblinear::SolverType::L1R_LR if @params[:penalty] == 'l1'
+        return Numo::Liblinear::SolverType::L2R_LR_DUAL if @params[:dual]
+
+        Numo::Liblinear::SolverType::L2R_LR
+      end
+
+      def binary_class?
+        @model[:nr_class] == 2
+      end
+
+      def fit_bias?
+        @params[:fit_bias]
+      end
+
+      def bias_scale
+        @params[:bias_scale]
+      end
+
+      def n_classes
+        @model[:nr_class]
+      end
+
+      def n_features
+        @model[:nr_feature]
+      end
+    end
+  end
+end

data/lib/rumale/svm/nu_svc.rb

@@ -0,0 +1,193 @@
+# frozen_string_literal: true
+
+require 'numo/libsvm'
+require 'rumale/base/base_estimator'
+require 'rumale/base/classifier'
+
+module Rumale
+  module SVM
+    # NuSVC is a class that provides Kernel Nu-Support Vector Classifier in LIBSVM with Rumale interface.
+    #
+    # @example
+    #   estimator = Rumale::SVM::NuSVC.new(nu: 0.5, kernel: 'rbf', gamma: 10.0, random_seed: 1)
+    #   estimator.fit(training_samples, traininig_labels)
+    #   results = estimator.predict(testing_samples)
+    class NuSVC
+      include Base::BaseEstimator
+      include Base::Classifier
+
+      # Create a new classifier with Kernel Nu-Support Vector Classifier.
+      #
+      # @param nu [Float] The regularization parameter. The interval of nu is (0, 1].
+      # @param kernel [String] The type of kernel function ('rbf', 'linear', 'poly', 'sigmoid', and 'precomputed').
+      # @param degree [Integer] The degree parameter in polynomial kernel function.
+      # @param gamma [Float] The gamma parameter in rbf/poly/sigmoid kernel function.
+      # @param coef0 [Float] The coefficient in poly/sigmoid kernel function.
+      # @param shrinking [Boolean] The flag indicating whether to use the shrinking heuristics.
+      # @param probability [Boolean] The flag indicating whether to train the parameter for probability estimation.
+      # @param cache_size [Float] The cache memory size in MB.
+      # @param tol [Float] The tolerance of termination criterion.
+      # @param verbose [Boolean] The flag indicating whether to output learning process message
+      # @param random_seed [Integer/Nil] The seed value using to initialize the random generator.
+      def initialize(nu: 0.5, kernel: 'rbf', degree: 3, gamma: 1.0, coef0: 0.0,
+                     shrinking: true, probability: true, cache_size: 200.0, tol: 1e-3, verbose: false, random_seed: nil)
+        check_params_float(nu: nu, gamma: gamma, coef0: coef0, cache_size: cache_size, tol: tol)
+        check_params_integer(degree: degree)
+        check_params_boolean(shrinking: shrinking, probability: probability, verbose: verbose)
+        check_params_type_or_nil(Integer, random_seed: random_seed)
+        @params = {}
+        @params[:nu] = nu
+        @params[:kernel] = kernel
+        @params[:degree] = degree
+        @params[:gamma] = gamma
+        @params[:coef0] = coef0
+        @params[:shrinking] = shrinking
+        @params[:probability] = probability
+        @params[:cache_size] = cache_size
+        @params[:tol] = tol
+        @params[:verbose] = verbose
+        @params[:random_seed] = random_seed
+        @model = nil
+      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.
+      #   If the kernel is 'precomputed', x must be a square distance matrix (shape: [n_samples, n_samples]).
+      # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
+      # @return [NuSVC] The learned classifier itself.
+      def fit(x, y)
+        check_sample_array(x)
+        check_label_array(y)
+        check_sample_label_size(x, y)
+        xx = precomputed_kernel? ? add_index_col(x) : x
+        @model = Numo::Libsvm.train(xx, y, libsvm_params)
+        self
+      end
+
+      # Calculate confidence scores for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
+      #   If the kernel is 'precomputed', the shape of x must be [n_samples, n_training_samples].
+      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Confidence score per sample.
+      def decision_function(x)
+        check_sample_array(x)
+        xx = precomputed_kernel? ? add_index_col(x) : x
+        Numo::Libsvm.decision_function(xx, libsvm_params, @model)
+      end
+
+      # Predict class labels for samples.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
+      #   If the kernel is 'precomputed', the shape of x must be [n_samples, n_training_samples].
+      # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
+      def predict(x)
+        check_sample_array(x)
+        xx = precomputed_kernel? ? add_index_col(x) : x
+        Numo::Int32.cast(Numo::Libsvm.predict(xx, libsvm_params, @model))
+      end
+
+      # Predict class probability for samples.
+      # This method works correctly only if the probability parameter is true.
+      #
+      # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
+      #   If the kernel is 'precomputed', the shape of x must be [n_samples, n_training_samples].
+      # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
+      def predict_proba(x)
+        check_sample_array(x)
+        xx = precomputed_kernel? ? add_index_col(x) : x
+        Numo::Libsvm.predict_proba(xx, libsvm_params, @model)
+      end
+
+      # Dump marshal data.
+      # @return [Hash] The marshal data about NuSVC.
+      def marshal_dump
+        { params: @params,
+          model: @model }
+      end
+
+      # Load marshal data.
+      # @return [nil]
+      def marshal_load(obj)
+        @params = obj[:params]
+        @model = obj[:model]
+        nil
+      end
+
+      # Return the indices of support vectors.
+      # @return [Numo::Int32] (shape: [n_support_vectors])
+      def support
+        @model[:sv_indices]
+      end
+
+      # Return the support_vectors.
+      # @return [Numo::DFloat] (shape: [n_support_vectors, n_features])
+      def support_vectors
+        precomputed_kernel? ? del_index_col(@model[:SV]) : @model[:SV]
+      end
+
+      # Return the number of support vectors for each class.
+      # @return [Numo::Int32] (shape: [n_classes])
+      def n_support
+        @model[:nSV]
+      end
+
+      # Return the coefficients of the support vector in decision function.
+      # @return [Numo::DFloat] (shape: [n_classes - 1, n_support_vectors])
+      def duel_coef
+        @model[:sv_coef]
+      end
+
+      # Return the intercepts in decision function.
+      # @return [Numo::DFloat] (shape: [n_classes * (n_classes - 1) / 2])
+      def intercept
+        @model[:rho]
+      end
+
+      # Return the probability parameter alpha.
+      # @return [Numo::DFloat] (shape: [n_classes * (n_classes - 1) / 2])
+      def prob_a
+        @model[:probA]
+      end
+
+      # Return the probability parameter beta.
+      # @return [Numo::DFloat] (shape: [n_classes * (n_classes - 1) / 2])
+      def prob_b
+        @model[:probB]
+      end
+
+      private
+
+      def add_index_col(x)
+        idx = Numo::Int32.new(x.shape[0]).seq + 1
+        Numo::NArray.hstack([idx.expand_dims(1), x])
+      end
+
+      def del_index_col(x)
+        x[true, 1..-1].dup
+      end
+
+      def precomputed_kernel?
+        @params[:kernel] == 'precomputed'
+      end
+
+      def libsvm_params
+        res = @params.merge(svm_type: Numo::Libsvm::SvmType::C_SVC)
+        res[:kernel_type] = case res.delete(:kernel)
+                            when 'linear'
+                              Numo::Libsvm::KernelType::LINEAR
+                            when 'poly'
+                              Numo::Libsvm::KernelType::POLY
+                            when 'sigmoid'
+                              Numo::Libsvm::KernelType::SIGMOID
+                            when 'precomputed'
+                              Numo::Libsvm::KernelType::PRECOMPUTED
+                            else
+                              Numo::Libsvm::KernelType::RBF
+                            end
+        res[:eps] = res.delete(:tol)
+        res
+      end
+    end
+  end
+end