rumale-linear_model 0.24.0 → 0.25.0

Sign up to get free protection for your applications and to get access to all the features.
@@ -0,0 +1,138 @@
1
+ # frozen_string_literal: true
2
+
3
+ require 'rumale/base/regressor'
4
+ require 'rumale/validation'
5
+
6
+ require_relative 'sgd_estimator'
7
+
8
+ module Rumale
9
+ module LinearModel
10
+ # SGDRegressor is a class that implements linear regressor with stochastic gradient descent optimization.
11
+ #
12
+ # @example
13
+ # require 'rumale/linear_model/sgd_regressor'
14
+ #
15
+ # estimator =
16
+ # Rumale::LinearModel::SGDRegressor.new(loss: 'squared_error', reg_param: 1.0, max_iter: 1000, batch_size: 50, random_seed: 1)
17
+ # estimator.fit(training_samples, traininig_target_values)
18
+ # results = estimator.predict(testing_samples)
19
+ #
20
+ # *Reference*
21
+ # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
22
+ # - 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.
23
+ # - Bottou, L., "Large-Scale Machine Learning with Stochastic Gradient Descent," Proc. COMPSTAT'10, pp. 177--186, 2010.
24
+ class SGDRegressor < Rumale::LinearModel::SGDEstimator
25
+ include Rumale::Base::Regressor
26
+
27
+ # Return the random generator for performing random sampling.
28
+ # @return [Random]
29
+ attr_reader :rng
30
+
31
+ # Create a new linear regressor with stochastic gradient descent optimization.
32
+ #
33
+ # @param loss [String] The loss function to be used ('squared_error' and 'epsilon_insensitive').
34
+ # @param learning_rate [Float] The initial value of learning rate.
35
+ # The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
36
+ # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
37
+ # If nil is given, the decay sets to 'reg_param * learning_rate'.
38
+ # @param momentum [Float] The momentum factor.
39
+ # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
40
+ # @param l1_ratio [Float] The elastic-net type regularization mixing parameter.
41
+ # If penalty set to 'l2' or 'l1', this parameter is ignored.
42
+ # If l1_ratio = 1, the regularization is similar to Lasso.
43
+ # If l1_ratio = 0, the regularization is similar to Ridge.
44
+ # If 0 < l1_ratio < 1, the regularization is a combination of L1 and L2.
45
+ # @param reg_param [Float] The regularization parameter.
46
+ # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
47
+ # @param bias_scale [Float] The scale of the bias term.
48
+ # @param epsilon [Float] The margin of tolerance. If loss set to 'squared_error', this parameter is ignored.
49
+ # @param max_iter [Integer] The maximum number of epochs that indicates
50
+ # how many times the whole data is given to the training process.
51
+ # @param batch_size [Integer] The size of the mini batches.
52
+ # @param tol [Float] The tolerance of loss for terminating optimization.
53
+ # @param n_jobs [Integer] The number of jobs for running the fit method in parallel.
54
+ # If nil is given, the method does not execute in parallel.
55
+ # If zero or less is given, it becomes equal to the number of processors.
56
+ # This parameter is ignored if the Parallel gem is not loaded.
57
+ # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
58
+ # @param random_seed [Integer] The seed value using to initialize the random generator.
59
+ def initialize(loss: 'squared_error', learning_rate: 0.01, decay: nil, momentum: 0.9,
60
+ penalty: 'l2', reg_param: 1.0, l1_ratio: 0.5,
61
+ fit_bias: true, bias_scale: 1.0,
62
+ epsilon: 0.1,
63
+ max_iter: 1000, batch_size: 50, tol: 1e-4,
64
+ n_jobs: nil, verbose: false, random_seed: nil)
65
+ super()
66
+ @params.merge!(
67
+ loss: loss,
68
+ learning_rate: learning_rate,
69
+ decay: decay,
70
+ momentum: momentum,
71
+ penalty: penalty,
72
+ reg_param: reg_param,
73
+ l1_ratio: l1_ratio,
74
+ fit_bias: fit_bias,
75
+ bias_scale: bias_scale,
76
+ epsilon: epsilon,
77
+ max_iter: max_iter,
78
+ batch_size: batch_size,
79
+ tol: tol,
80
+ n_jobs: n_jobs,
81
+ verbose: verbose,
82
+ random_seed: random_seed
83
+ )
84
+ @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
85
+ @params[:random_seed] ||= srand
86
+ @rng = Random.new(@params[:random_seed])
87
+ @penalty_type = @params[:penalty]
88
+ @loss_func = case @params[:loss]
89
+ when Rumale::LinearModel::Loss::MeanSquaredError::NAME
90
+ Rumale::LinearModel::Loss::MeanSquaredError.new
91
+ when Rumale::LinearModel::Loss::EpsilonInsensitive::NAME
92
+ Rumale::LinearModel::Loss::EpsilonInsensitive.new(epsilon: @params[:epsilon])
93
+ else
94
+ raise ArgumentError, "given loss '#{loss}' is not supported."
95
+ end
96
+ end
97
+
98
+ # Fit the model with given training data.
99
+ #
100
+ # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The training data to be used for fitting the model.
101
+ # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
102
+ # @retu:rn [SGDRegressor] The learned regressor itself.
103
+ def fit(x, y)
104
+ x = Rumale::Validation.check_convert_sample_array(x)
105
+ y = Rumale::Validation.check_convert_target_value_array(y)
106
+ Rumale::Validation.check_sample_size(x, y)
107
+
108
+ n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
109
+ n_features = x.shape[1]
110
+
111
+ if n_outputs > 1
112
+ @weight_vec = Numo::DFloat.zeros(n_outputs, n_features)
113
+ @bias_term = Numo::DFloat.zeros(n_outputs)
114
+ if enable_parallel?
115
+ models = parallel_map(n_outputs) { |n| partial_fit(x, y[true, n]) }
116
+ n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = models[n] }
117
+ else
118
+ n_outputs.times { |n| @weight_vec[n, true], @bias_term[n] = partial_fit(x, y[true, n]) }
119
+ end
120
+ else
121
+ @weight_vec, @bias_term = partial_fit(x, y)
122
+ end
123
+
124
+ self
125
+ end
126
+
127
+ # Predict values for samples.
128
+ #
129
+ # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the values.
130
+ # @return [Numo::DFloat] (shape: [n_samples, n_outputs]) Predicted values per sample.
131
+ def predict(x)
132
+ x = Rumale::Validation.check_convert_sample_array(x)
133
+
134
+ x.dot(@weight_vec.transpose) + @bias_term
135
+ end
136
+ end
137
+ end
138
+ end
@@ -1,15 +1,16 @@
1
1
  # frozen_string_literal: true
2
2
 
3
+ require 'lbfgsb'
4
+
3
5
  require 'rumale/base/classifier'
4
- require 'rumale/linear_model/base_sgd'
5
6
  require 'rumale/probabilistic_output'
6
7
  require 'rumale/validation'
7
8
 
9
+ require_relative 'base_estimator'
10
+
8
11
  module Rumale
9
- # This module consists of the classes that implement generalized linear models.
10
12
  module LinearModel
11
- # SVC is a class that implements Support Vector Classifier
12
- # with stochastic gradient descent optimization.
13
+ # SVC is a class that implements Support Vector Classifier with the squared hinge loss.
13
14
  # For multiclass classification problem, it uses one-vs-the-rest strategy.
14
15
  #
15
16
  # @note
@@ -21,52 +22,23 @@ module Rumale
21
22
  # require 'rumale/linear_model/svc'
22
23
  #
23
24
  # estimator =
24
- # Rumale::LinearModel::SVC.new(reg_param: 1.0, max_iter: 1000, batch_size: 50, random_seed: 1)
25
+ # Rumale::LinearModel::SVC.new(reg_param: 1.0)
25
26
  # estimator.fit(training_samples, traininig_labels)
26
27
  # results = estimator.predict(testing_samples)
27
- #
28
- # *Reference*
29
- # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
30
- # - 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.
31
- # - Bottou, L., "Large-Scale Machine Learning with Stochastic Gradient Descent," Proc. COMPSTAT'10, pp. 177--186, 2010.
32
- class SVC < BaseSGD
33
- include ::Rumale::Base::Classifier
34
-
35
- # Return the weight vector for SVC.
36
- # @return [Numo::DFloat] (shape: [n_classes, n_features])
37
- attr_reader :weight_vec
38
-
39
- # Return the bias term (a.k.a. intercept) for SVC.
40
- # @return [Numo::DFloat] (shape: [n_classes])
41
- attr_reader :bias_term
28
+ class SVC < Rumale::LinearModel::BaseEstimator
29
+ include Rumale::Base::Classifier
42
30
 
43
31
  # Return the class labels.
44
32
  # @return [Numo::Int32] (shape: [n_classes])
45
33
  attr_reader :classes
46
34
 
47
- # Return the random generator for performing random sampling.
48
- # @return [Random]
49
- attr_reader :rng
50
-
51
- # Create a new classifier with Support Vector Machine by the SGD optimization.
35
+ # Create a new linear classifier with Support Vector Machine with the squared hinge loss.
52
36
  #
53
- # @param learning_rate [Float] The initial value of learning rate.
54
- # The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
55
- # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
56
- # If nil is given, the decay sets to 'reg_param * learning_rate'.
57
- # @param momentum [Float] The momentum factor.
58
- # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
59
- # @param l1_ratio [Float] The elastic-net type regularization mixing parameter.
60
- # If penalty set to 'l2' or 'l1', this parameter is ignored.
61
- # If l1_ratio = 1, the regularization is similar to Lasso.
62
- # If l1_ratio = 0, the regularization is similar to Ridge.
63
- # If 0 < l1_ratio < 1, the regularization is a combination of L1 and L2.
64
37
  # @param reg_param [Float] The regularization parameter.
65
38
  # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
66
39
  # @param bias_scale [Float] The scale of the bias term.
67
40
  # @param max_iter [Integer] The maximum number of epochs that indicates
68
41
  # how many times the whole data is given to the training process.
69
- # @param batch_size [Integer] The size of the mini batches.
70
42
  # @param tol [Float] The tolerance of loss for terminating optimization.
71
43
  # @param probability [Boolean] The flag indicating whether to perform probability estimation.
72
44
  # @param n_jobs [Integer] The number of jobs for running the fit and predict methods in parallel.
@@ -74,20 +46,20 @@ module Rumale
74
46
  # If zero or less is given, it becomes equal to the number of processors.
75
47
  # This parameter is ignored if the Parallel gem is not loaded.
76
48
  # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
77
- # @param random_seed [Integer] The seed value using to initialize the random generator.
78
- def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
79
- penalty: 'l2', reg_param: 1.0, l1_ratio: 0.5,
80
- fit_bias: true, bias_scale: 1.0,
81
- max_iter: 1000, batch_size: 50, tol: 1e-4,
82
- probability: false,
83
- n_jobs: nil, verbose: false, random_seed: nil)
49
+ # 'iterate.dat' file is generated by lbfgsb.rb.
50
+ def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0, max_iter: 1000, tol: 1e-4, probability: false,
51
+ n_jobs: nil, verbose: false)
84
52
  super()
85
- @params.merge!(method(:initialize).parameters.to_h { |_t, arg| [arg, binding.local_variable_get(arg)] })
86
- @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
87
- @params[:random_seed] ||= srand
88
- @rng = Random.new(@params[:random_seed])
89
- @penalty_type = @params[:penalty]
90
- @loss_func = ::Rumale::LinearModel::Loss::HingeLoss.new
53
+ @params = {
54
+ reg_param: reg_param,
55
+ fit_bias: fit_bias,
56
+ bias_scale: bias_scale,
57
+ max_iter: max_iter,
58
+ tol: tol,
59
+ probability: probability,
60
+ n_jobs: n_jobs,
61
+ verbose: verbose
62
+ }
91
63
  end
92
64
 
93
65
  # Fit the model with given training data.
@@ -96,20 +68,20 @@ module Rumale
96
68
  # @param y [Numo::Int32] (shape: [n_samples]) The labels to be used for fitting the model.
97
69
  # @return [SVC] The learned classifier itself.
98
70
  def fit(x, y)
99
- x = ::Rumale::Validation.check_convert_sample_array(x)
100
- y = ::Rumale::Validation.check_convert_label_array(y)
101
- ::Rumale::Validation.check_sample_size(x, y)
71
+ x = Rumale::Validation.check_convert_sample_array(x)
72
+ y = Rumale::Validation.check_convert_label_array(y)
73
+ Rumale::Validation.check_sample_size(x, y)
102
74
 
103
75
  @classes = Numo::Int32[*y.to_a.uniq.sort]
76
+ x = expand_feature(x) if fit_bias?
104
77
 
105
78
  if multiclass_problem?
106
79
  n_classes = @classes.size
107
80
  n_features = x.shape[1]
108
- # initialize model.
81
+ n_features -= 1 if fit_bias?
109
82
  @weight_vec = Numo::DFloat.zeros(n_classes, n_features)
110
83
  @bias_term = Numo::DFloat.zeros(n_classes)
111
84
  @prob_param = Numo::DFloat.zeros(n_classes, 2)
112
- # fit model.
113
85
  models = if enable_parallel?
114
86
  parallel_map(n_classes) do |n|
115
87
  bin_y = Numo::Int32.cast(y.eq(@classes[n])) * 2 - 1
@@ -121,7 +93,6 @@ module Rumale
121
93
  partial_fit(x, bin_y)
122
94
  end
123
95
  end
124
- # store model.
125
96
  models.each_with_index { |model, n| @weight_vec[n, true], @bias_term[n], @prob_param[n, true] = model }
126
97
  else
127
98
  negative_label = @classes[0]
@@ -137,7 +108,7 @@ module Rumale
137
108
  # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to compute the scores.
138
109
  # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Confidence score per sample.
139
110
  def decision_function(x)
140
- x = ::Rumale::Validation.check_convert_sample_array(x)
111
+ x = Rumale::Validation.check_convert_sample_array(x)
141
112
 
142
113
  x.dot(@weight_vec.transpose) + @bias_term
143
114
  end
@@ -147,7 +118,7 @@ module Rumale
147
118
  # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the labels.
148
119
  # @return [Numo::Int32] (shape: [n_samples]) Predicted class label per sample.
149
120
  def predict(x)
150
- x = ::Rumale::Validation.check_convert_sample_array(x)
121
+ x = Rumale::Validation.check_convert_sample_array(x)
151
122
 
152
123
  n_samples = x.shape[0]
153
124
  predicted = if multiclass_problem?
@@ -169,13 +140,13 @@ module Rumale
169
140
  # @param x [Numo::DFloat] (shape: [n_samples, n_features]) The samples to predict the probailities.
170
141
  # @return [Numo::DFloat] (shape: [n_samples, n_classes]) Predicted probability of each class per sample.
171
142
  def predict_proba(x)
172
- x = ::Rumale::Validation.check_convert_sample_array(x)
143
+ x = Rumale::Validation.check_convert_sample_array(x)
173
144
 
174
145
  if multiclass_problem?
175
146
  probs = 1.0 / (Numo::NMath.exp(@prob_param[true, 0] * decision_function(x) + @prob_param[true, 1]) + 1.0)
176
147
  (probs.transpose / probs.sum(axis: 1)).transpose.dup
177
148
  else
178
- n_samples, = x.shape
149
+ n_samples = x.shape[0]
179
150
  probs = Numo::DFloat.zeros(n_samples, 2)
180
151
  probs[true, 1] = 1.0 / (Numo::NMath.exp(@prob_param[0] * decision_function(x) + @prob_param[1]) + 1.0)
181
152
  probs[true, 0] = 1.0 - probs[true, 1]
@@ -185,14 +156,35 @@ module Rumale
185
156
 
186
157
  private
187
158
 
188
- def partial_fit(x, bin_y)
189
- w, b = super
190
- p = if @params[:probability]
191
- ::Rumale::ProbabilisticOutput.fit_sigmoid(x.dot(w.transpose) + b, bin_y)
192
- else
193
- Numo::DFloat[1, 0]
194
- end
195
- [w, b, p]
159
+ def partial_fit(base_x, bin_y)
160
+ fnc = proc do |w, x, y, reg_param|
161
+ n_samples = x.shape[0]
162
+ z = x.dot(w)
163
+ t = 1 - y * z
164
+ loss = 0.5 * reg_param * w.dot(w) + (x.class.maximum(0, t)**2).sum.fdiv(n_samples)
165
+ indices = t.gt(0)
166
+ grad = reg_param * w
167
+ if indices.count.positive?
168
+ sx = x[indices, true]
169
+ sy = y[indices]
170
+ grad += 2.fdiv(n_samples) * sx.transpose.dot((sx.dot(w) - sy))
171
+ end
172
+ [loss, grad]
173
+ end
174
+
175
+ n_features = base_x.shape[1]
176
+ w_init = Numo::DFloat.zeros(n_features)
177
+
178
+ res = Lbfgsb.minimize(
179
+ fnc: fnc, jcb: true, x_init: w_init, args: [base_x, bin_y, @params[:reg_param]],
180
+ maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
181
+ verbose: @params[:verbose] ? 1 : -1
182
+ )
183
+
184
+ prb = @params[:probability] ? Rumale::ProbabilisticOutput.fit_sigmoid(base_x.dot(res[:x]), bin_y) : Numo::DFloat[1, 0]
185
+ w, b = split_weight(res[:x])
186
+
187
+ [w, b, prb]
196
188
  end
197
189
 
198
190
  def multiclass_problem?
@@ -1,13 +1,15 @@
1
1
  # frozen_string_literal: true
2
2
 
3
+ require 'lbfgsb'
4
+
3
5
  require 'rumale/base/regressor'
4
6
  require 'rumale/validation'
5
- require 'rumale/linear_model/base_sgd'
7
+
8
+ require_relative 'base_estimator'
6
9
 
7
10
  module Rumale
8
11
  module LinearModel
9
- # SVR is a class that implements Support Vector Regressor
10
- # with stochastic gradient descent optimization.
12
+ # SVR is a class that implements Support Vector Regressor with the squared epsilon-insensitive loss.
11
13
  #
12
14
  # @note
13
15
  # Rumale::SVM provides linear and kernel support vector regressor based on LIBLINEAR and LIBSVM.
@@ -17,70 +19,39 @@ module Rumale
17
19
  # @example
18
20
  # require 'rumale/linear_model/svr'
19
21
  #
20
- # estimator =
21
- # Rumale::LinearModel::SVR.new(reg_param: 1.0, epsilon: 0.1, max_iter: 1000, batch_size: 50, random_seed: 1)
22
+ # estimator = Rumale::LinearModel::SVR.new(reg_param: 1.0, epsilon: 0.1)
22
23
  # estimator.fit(training_samples, traininig_target_values)
23
24
  # results = estimator.predict(testing_samples)
24
- #
25
- # *Reference*
26
- # - Shalev-Shwartz, S., and Singer, Y., "Pegasos: Primal Estimated sub-GrAdient SOlver for SVM," Proc. ICML'07, pp. 807--814, 2007.
27
- # - 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.
28
- # - Bottou, L., "Large-Scale Machine Learning with Stochastic Gradient Descent," Proc. COMPSTAT'10, pp. 177--186, 2010.
29
- class SVR < BaseSGD
30
- include ::Rumale::Base::Regressor
31
-
32
- # Return the weight vector for SVR.
33
- # @return [Numo::DFloat] (shape: [n_outputs, n_features])
34
- attr_reader :weight_vec
35
-
36
- # Return the bias term (a.k.a. intercept) for SVR.
37
- # @return [Numo::DFloat] (shape: [n_outputs])
38
- attr_reader :bias_term
39
-
40
- # Return the random generator for performing random sampling.
41
- # @return [Random]
42
- attr_reader :rng
25
+ class SVR < Rumale::LinearModel::BaseEstimator
26
+ include Rumale::Base::Regressor
43
27
 
44
28
  # Create a new regressor with Support Vector Machine by the SGD optimization.
45
29
  #
46
- # @param learning_rate [Float] The initial value of learning rate.
47
- # The learning rate decreases as the iteration proceeds according to the equation: learning_rate / (1 + decay * t).
48
- # @param decay [Float] The smoothing parameter for decreasing learning rate as the iteration proceeds.
49
- # If nil is given, the decay sets to 'reg_param * learning_rate'.
50
- # @param momentum [Float] The momentum factor.
51
- # @param penalty [String] The regularization type to be used ('l1', 'l2', and 'elasticnet').
52
- # @param l1_ratio [Float] The elastic-net type regularization mixing parameter.
53
- # If penalty set to 'l2' or 'l1', this parameter is ignored.
54
- # If l1_ratio = 1, the regularization is similar to Lasso.
55
- # If l1_ratio = 0, the regularization is similar to Ridge.
56
- # If 0 < l1_ratio < 1, the regularization is a combination of L1 and L2.
57
30
  # @param reg_param [Float] The regularization parameter.
58
31
  # @param fit_bias [Boolean] The flag indicating whether to fit the bias term.
59
32
  # @param bias_scale [Float] The scale of the bias term.
60
33
  # @param epsilon [Float] The margin of tolerance.
61
34
  # @param max_iter [Integer] The maximum number of epochs that indicates
62
35
  # how many times the whole data is given to the training process.
63
- # @param batch_size [Integer] The size of the mini batches.
64
36
  # @param tol [Float] The tolerance of loss for terminating optimization.
65
37
  # @param n_jobs [Integer] The number of jobs for running the fit method in parallel.
66
38
  # If nil is given, the method does not execute in parallel.
67
39
  # If zero or less is given, it becomes equal to the number of processors.
68
40
  # This parameter is ignored if the Parallel gem is not loaded.
69
41
  # @param verbose [Boolean] The flag indicating whether to output loss during iteration.
70
- # @param random_seed [Integer] The seed value using to initialize the random generator.
71
- def initialize(learning_rate: 0.01, decay: nil, momentum: 0.9,
72
- penalty: 'l2', reg_param: 1.0, l1_ratio: 0.5,
73
- fit_bias: true, bias_scale: 1.0,
74
- epsilon: 0.1,
75
- max_iter: 1000, batch_size: 50, tol: 1e-4,
76
- n_jobs: nil, verbose: false, random_seed: nil)
42
+ def initialize(reg_param: 1.0, fit_bias: true, bias_scale: 1.0, epsilon: 0.1, max_iter: 1000, tol: 1e-4,
43
+ n_jobs: nil, verbose: false)
77
44
  super()
78
- @params.merge!(method(:initialize).parameters.to_h { |_t, arg| [arg, binding.local_variable_get(arg)] })
79
- @params[:decay] ||= @params[:reg_param] * @params[:learning_rate]
80
- @params[:random_seed] ||= srand
81
- @rng = Random.new(@params[:random_seed])
82
- @penalty_type = @params[:penalty]
83
- @loss_func = ::Rumale::LinearModel::Loss::EpsilonInsensitive.new(epsilon: @params[:epsilon])
45
+ @params = {
46
+ reg_param: reg_param,
47
+ fit_bias: fit_bias,
48
+ bias_scale: bias_scale,
49
+ epsilon: epsilon,
50
+ max_iter: max_iter,
51
+ tol: tol,
52
+ n_jobs: n_jobs,
53
+ verbose: verbose
54
+ }
84
55
  end
85
56
 
86
57
  # Fit the model with given training data.
@@ -89,9 +60,9 @@ module Rumale
89
60
  # @param y [Numo::DFloat] (shape: [n_samples, n_outputs]) The target values to be used for fitting the model.
90
61
  # @return [SVR] The learned regressor itself.
91
62
  def fit(x, y)
92
- x = ::Rumale::Validation.check_convert_sample_array(x)
93
- y = ::Rumale::Validation.check_convert_target_value_array(y)
94
- ::Rumale::Validation.check_sample_size(x, y)
63
+ x = Rumale::Validation.check_convert_sample_array(x)
64
+ y = Rumale::Validation.check_convert_target_value_array(y)
65
+ Rumale::Validation.check_sample_size(x, y)
95
66
 
96
67
  n_outputs = y.shape[1].nil? ? 1 : y.shape[1]
97
68
  n_features = x.shape[1]
@@ -121,6 +92,37 @@ module Rumale
121
92
 
122
93
  x.dot(@weight_vec.transpose) + @bias_term
123
94
  end
95
+
96
+ private
97
+
98
+ def partial_fit(base_x, single_y)
99
+ fnc = proc do |w, x, y, eps, reg_param|
100
+ n_samples = x.shape[0]
101
+ z = x.dot(w)
102
+ d = y - z
103
+ loss = 0.5 * reg_param * w.dot(w) + (x.class.maximum(0, d.abs - eps)**2).sum.fdiv(n_samples)
104
+ c = x.class.zeros(n_samples)
105
+ indices = d.gt(eps)
106
+ c[indices] = -d[indices] + eps if indices.count.positive?
107
+ indices = d.lt(eps)
108
+ c[indices] = -d[indices] - eps if indices.count.positive?
109
+ grad = reg_param * w + 2.fdiv(n_samples) * x.transpose.dot(c)
110
+ [loss, grad]
111
+ end
112
+
113
+ base_x = expand_feature(base_x) if fit_bias?
114
+
115
+ n_features = base_x.shape[1]
116
+ w_init = Numo::DFloat.zeros(n_features)
117
+
118
+ res = Lbfgsb.minimize(
119
+ fnc: fnc, jcb: true, x_init: w_init, args: [base_x, single_y, @params[:epsilon], @params[:reg_param]],
120
+ maxiter: @params[:max_iter], factr: @params[:tol] / Lbfgsb::DBL_EPSILON,
121
+ verbose: @params[:verbose] ? 1 : -1
122
+ )
123
+
124
+ split_weight(res[:x])
125
+ end
124
126
  end
125
127
  end
126
128
  end
@@ -5,6 +5,6 @@ module Rumale
5
5
  # This module consists of the classes that implement generalized linear models.
6
6
  module LinearModel
7
7
  # @!visibility private
8
- VERSION = '0.24.0'
8
+ VERSION = '0.25.0'
9
9
  end
10
10
  end
@@ -2,7 +2,8 @@
2
2
 
3
3
  require 'numo/narray'
4
4
 
5
- require_relative 'linear_model/base_sgd'
5
+ require_relative 'linear_model/sgd_classifier'
6
+ require_relative 'linear_model/sgd_regressor'
6
7
  require_relative 'linear_model/elastic_net'
7
8
  require_relative 'linear_model/lasso'
8
9
  require_relative 'linear_model/linear_regression'
metadata CHANGED
@@ -1,14 +1,14 @@
1
1
  --- !ruby/object:Gem::Specification
2
2
  name: rumale-linear_model
3
3
  version: !ruby/object:Gem::Version
4
- version: 0.24.0
4
+ version: 0.25.0
5
5
  platform: ruby
6
6
  authors:
7
7
  - yoshoku
8
8
  autorequire:
9
9
  bindir: exe
10
10
  cert_chain: []
11
- date: 2022-12-31 00:00:00.000000000 Z
11
+ date: 2023-01-18 00:00:00.000000000 Z
12
12
  dependencies:
13
13
  - !ruby/object:Gem::Dependency
14
14
  name: lbfgsb
@@ -44,14 +44,14 @@ dependencies:
44
44
  requirements:
45
45
  - - "~>"
46
46
  - !ruby/object:Gem::Version
47
- version: 0.24.0
47
+ version: 0.25.0
48
48
  type: :runtime
49
49
  prerelease: false
50
50
  version_requirements: !ruby/object:Gem::Requirement
51
51
  requirements:
52
52
  - - "~>"
53
53
  - !ruby/object:Gem::Version
54
- version: 0.24.0
54
+ version: 0.25.0
55
55
  description: |
56
56
  Rumale::LinearModel provides linear model algorithms,
57
57
  such as Logistic Regression, Support Vector Machine, Lasso, and Ridge Regression
@@ -65,13 +65,16 @@ files:
65
65
  - LICENSE.txt
66
66
  - README.md
67
67
  - lib/rumale/linear_model.rb
68
- - lib/rumale/linear_model/base_sgd.rb
68
+ - lib/rumale/linear_model/base_estimator.rb
69
69
  - lib/rumale/linear_model/elastic_net.rb
70
70
  - lib/rumale/linear_model/lasso.rb
71
71
  - lib/rumale/linear_model/linear_regression.rb
72
72
  - lib/rumale/linear_model/logistic_regression.rb
73
73
  - lib/rumale/linear_model/nnls.rb
74
74
  - lib/rumale/linear_model/ridge.rb
75
+ - lib/rumale/linear_model/sgd_classifier.rb
76
+ - lib/rumale/linear_model/sgd_estimator.rb
77
+ - lib/rumale/linear_model/sgd_regressor.rb
75
78
  - lib/rumale/linear_model/svc.rb
76
79
  - lib/rumale/linear_model/svr.rb
77
80
  - lib/rumale/linear_model/version.rb