glm 0.0.1 → 0.0.2

data/examples/test_glm_linear.rb

@@ -8,16 +8,17 @@ T = 2000
 
 iris = load_iris
 y_arr = iris.map {|r| r[0]}
+y = GSL::Vector.alloc y_arr
 x = iris.map {|r| r[1...-1]}
 x = x.map{|r| r << 1}
-x,y = GLM::Util.formatArrays(x, y_arr)
+x = GLM::Util.aa_to_gsl_matrix x
 
-linear = GLM::Linear.new(x, y)
+linear = GLM::Linear.new(x, y,0.05)
 
 (1..T).each do |i|
   linear.sto_update
   est_y =  linear.est(x)
-  err = (y.zip(est_y).map{|e| (e[0]-e[1])**2}).reduce(:+)/y.row_size
+  err = (est_y - y).nrm2
   puts err
 end
 

data/examples/test_glm_logit.rb

@@ -8,19 +8,17 @@ T = 200
 
 iris = load_iris
 y_arr = iris.map {|r| r[-1].to_i}
+y = GSL::Vector.alloc y_arr
 x = iris.map {|r| r[0...-2]}
 x = x.map{|r| r << 1}
-x,y = GLM::Util.formatArrays(x, y_arr)
+x = GLM::Util.aa_to_gsl_matrix x
 
 glm = GLM::Logit.new(x, y, 0.1)
 
 (1..T).each do |i|
   glm.sto_update
-
   est_y =  glm.est(x)
-  err = (y.zip(est_y).map{|e| (e[0]!=e[1])?1:0}).reduce(:+)#/y.row_size
-#  pp y.zip(est_y)
-  
+  err = (est_y - y).nrm2
   puts err
   pp glm.theta
 end

data/lib/glm.rb

@@ -1,4 +1,5 @@
 require 'matrix'
+require 'gsl'
 include Math
 
 

data/lib/glm/base.rb

@@ -1,10 +1,10 @@
 class GLM::Base
-
-  def initialize(x,y,alpha = 0.1)
+  @@initial_weight = 1
+  def initialize(x,y,alpha = 0.05)
     @x = x
     @y = y
     @@alpha = alpha
-    @theta = Array.new(x.column_size,1)
+    @theta = GSL::Vector.alloc(Array.new(x.size2, @@initial_weight))
   end
 
   #Log partition function <b>a(eta)</b>, intended to be overriden
@@ -18,22 +18,16 @@ class GLM::Base
   end
 
   def format(x)
-    if x.is_a? Array
-      if x[0].is_a?(Array)
-        x.map {|e|
-          output(e)}
-      else
-        output(x)
-      end
-    #Assuming x.is_a?(Matrix) == true
-    else
-      x.row_vectors.map {|e|
-        output(Matrix.row_vector(e))
-      }
+    if x.is_a? GSL::Vector
+      return output(x)
+    elsif x.is_a? GSL::Matrix
+      tmp = GSL::Vector.alloc x.size1
+      (0...x.size1).each {|i|
+        tmp[i]= output(x.row(i))}
+      return tmp
     end
   end
 
-
   # Estimator
   # =Arguments:
   #   x: a feature vector in Array
@@ -46,12 +40,12 @@ class GLM::Base
   #Output estimation from E(y|theta,x)
   #Need overriding, except for plain linear regression
   def output(x)
-    return h(x.t)
+    return h(x)
   end
 
   #Natural parameter eta
   def eta(x)
-    tmp = (Matrix.column_vector(@theta).t * x)[0,0]
+    tmp = @theta * x.transpose
     return tmp
   end
 
@@ -69,7 +63,8 @@ class GLM::Base
   #Gradient on one sample
   def gradient(x,y,v)
     tmp = h(v)
-    return (y - tmp) * x
+    res = (y - tmp) * x
+    return res
   end
 
   # Hypothesis function, outputs E(y|theta, x), mean of y given x parameterized by theta
@@ -89,11 +84,13 @@ class GLM::Base
 
   #One complete loop of stochastic gradient descend
   def sto_update()
-    (0...(@x.row_size)).each do |i|
-      (0...(@x.column_size)).each do |j|
-        @theta[j] += @@alpha * gradient(@x[i,j], @y[i,0], Matrix.column_vector(@x.row(i)))
+    (0...(@x.size1)).each do |i|
+      (0...(@x.size2)).each do |j|
+        updates = gradient(@x[i,j], @y[i], @x.row(i))
+        @theta[j] = @theta[j] + @@alpha * updates
       end
     end
+    pp @theta
   end
 
   def theta()

data/lib/glm/linear.rb

@@ -13,15 +13,16 @@ class GLM::Linear < GLM::Base
   def self.g(eta)
     return eta
   end
-
+  
+  #Normal equation estimation
   def ne_est(x)
     @theta = ne_fit
-    [@theta, x * Matrix.column_vector(@theta)]
+    [@theta, x * @theta]
   end
 
   #Normal equation fit
   def ne_fit
-    (( @x.t * @x ).inverse * @x.t * @y).to_a.flatten
+    ( @x.transpose * @x ).invert * @x.transpose * @y
   end
 
 end

data/lib/glm/logit.rb

@@ -17,7 +17,7 @@ class GLM::Logit < GLM::Base
   end
   
   def output(x)
-    return (h(x.t) > 0.5)?1:0
+    return (h(x) > 0.5)?1:0
   end
   
   def phi

data/lib/glm/util.rb

@@ -6,4 +6,16 @@ class GLM::Util
     y = Matrix.column_vector(y)
     return [x, y]
   end
+
+  #Construct a GSL::Matrix out of an Array of Arrays, should have been in GSL::Matrix already
+  def self.aa_to_gsl_matrix(aa)
+    n_rows = aa.length
+    n_cols = aa[0].length
+    gmat = GSL::Matrix.alloc n_rows,n_cols
+    (0...n_rows).each {|i|
+      (0...n_cols).each {|j|
+        gmat.set i, j, aa[i][j]}}
+    return gmat
+  end
+
 end

data/test/test_glm.rb

@@ -1,7 +1,9 @@
 require 'test/unit'
 require 'glm'
 require 'pp'
-require 'ruby-debug' ; Debugger.start(:post_mortem => true)
+require 'ruby-debug'
+Debugger.start(:post_mortem => true)
+Debugger.settings[:autoeval] = true
 
 class GLMTest < Test::Unit::TestCase
   def test_glm_logit_truth
@@ -30,10 +32,12 @@ class GLMTest < Test::Unit::TestCase
     y = iris.map {|r| r[0]}
     x = iris.map {|r| r[1...-1]}
     x = x.map{|r| r << 1}
-    x,y = GLM::Util.formatArrays(x, y)
+    #x,y = GLM::Util.formatArrays(x, y)
+    x = GLM::Util.aa_to_gsl_matrix(x)
+    y = GSL::Vector.alloc(y)
     linear = GLM::Linear.new(x, y)
     theta, y_est = linear.ne_est(x)
-    assert  ((y_est - y).map {|e| e ** 2}).reduce(:+)/y.row_size < 0.1
+    assert  ((y_est - y).map {|e| e ** 2}).norm < 0.1
   end
 
   

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: glm
 version: !ruby/object:Gem::Version
-  version: 0.0.1
+  version: 0.0.2
   prerelease: 
 platform: ruby
 authors:
@@ -48,8 +48,9 @@ required_rubygems_version: !ruby/object:Gem::Requirement
       version: '0'
 requirements: []
 rubyforge_project: 
-rubygems_version: 1.8.10
+rubygems_version: 1.8.19
 signing_key: 
 specification_version: 3
 summary: This is a W.I.P implementation of GLM
 test_files: []
+has_rdoc: