From 2b1d7f881c19c4d4b5afae194e02d3300c7675d0 Mon Sep 17 00:00:00 2001
From: Patrick Simianer <p@simianer.de>
Date: Tue, 5 Jul 2016 11:01:46 +0200
Subject: mv

---
 lin-reg | 70 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 70 insertions(+)
 create mode 100755 lin-reg

(limited to 'lin-reg')

diff --git a/lin-reg b/lin-reg
new file mode 100755
index 0000000..7a8e614
--- /dev/null
+++ b/lin-reg
@@ -0,0 +1,70 @@
+#!/usr/bin/env ruby
+
+require 'zipf'
+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
+  conf = Trollop::options do
+    opt :input,         "input data",         :type => :string, :required => true
+    opt :output,        "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 conf[:input], conf[:scale_features]
+  zeros = [0.0]*data[0].size
+  t = ReadFile.readlines(conf[: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|
+      loss = model.dot(x) - t[j]
+      overall_loss += loss**2
+      u += x * loss
+    }
+    STDERR.write "#{i} #{overall_loss/data.size}\n" if conf[:show_loss]
+    u *= conf[:learning_rate]*(1.0/t.size)
+    model -= u
+    if model.approx_eql? prev_model
+      stop += 1
+    else
+      stop = 0
+    end
+    break if stop==conf[:stop]
+    prev_model = model
+  end
+  tss = t.map{ |y| (y-t.mean)**2 }.sum
+  j = -1
+  rss = t.map{ |y| j+=1; (y-model.dot(data[j]))**2 }.sum
+  STDERR.write "ran for #{i} iterations\n R^2=#{1-(rss/tss)}\n"
+  puts model.to_s
+end
+
+main
+
-- 
cgit v1.2.3