From 26c490f404731d053a6205719b6246502c07b449 Mon Sep 17 00:00:00 2001
From: Patrick Simianer <p@simianer.de>
Date: Sat, 14 Jun 2014 16:46:27 +0200
Subject: init

---
 algorithms/log_reg_sgd.rb | 71 +++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 71 insertions(+)
 create mode 100755 algorithms/log_reg_sgd.rb

(limited to 'algorithms/log_reg_sgd.rb')

diff --git a/algorithms/log_reg_sgd.rb b/algorithms/log_reg_sgd.rb
new file mode 100755
index 0000000..16a48e2
--- /dev/null
+++ b/algorithms/log_reg_sgd.rb
@@ -0,0 +1,71 @@
+#!/usr/bin/env ruby
+
+require 'nlp_ruby'
+require 'trollop'
+
+
+def read_data fn, scale
+  f = ReadFile.new fn
+  data = []
+  while line = f.gets
+    line.strip!
+    a = []
+    a << 1.0
+    tokenize(line).each { |i| a << i.to_f }
+    v = SparseVector.from_a a
+    data << v
+  end
+  if scale
+    data.map { |i| i.keys }.flatten.uniq.each { |k|
+      max = data.map { |i| i[k] }.max
+      data.each { |i| i[k] /= max }
+    }
+  end
+  return data
+end
+
+def main
+  cfg = Trollop::options do
+    opt :input,         "input data",         :type => :string, :required => true
+    opt :output,        "1/0 output data",        :type => :string, :required => true
+    opt :learning_rate, "learning rate",      :type => :float,  :default => 0.07
+    opt :stop,          "stopping criterion", :type => :int,    :default => 100
+    opt :scale_features,"scale features",     :type => :bool,   :default => false, :short => '-t'
+    opt :show_loss,     "show loss per iter", :type => :bool,   :default => false
+  end
+  data = read_data cfg[:input], cfg[:scale_features]
+  zeros = [0.0]*data[0].size
+  t = ReadFile.readlines(cfg[:output]).map{ |i| i.to_f }
+  model = SparseVector.new zeros
+  stop = 0
+  prev_model = nil
+  i = 0
+  while true
+    i += 1
+    u = SparseVector.new zeros
+    overall_loss = 0.0
+    data.each_with_index { |x,j|
+      m = 1.0/(1+Math.exp(-model.dot(x)))
+      loss = m - t[j]
+      overall_loss += loss
+      u += x * loss
+    }
+    STDERR.write "#{i} #{overall_loss/data.size}\n" if cfg[:show_loss]
+    u *= cfg[:learning_rate]*(1.0/t.size)
+    model -= u
+    if model.approx_eql? prev_model
+      stop += 1
+    else
+      stop = 0
+    end
+    break if stop==cfg[:stop]
+    prev_model = model
+    puts model.to_s
+  end
+  STDERR.write "ran for #{i} iterations\n"
+  puts model.to_s
+end
+
+
+main
+
-- 
cgit v1.2.3