Added tool to optimize rnn with pybrain evolution.

Author: bayerj

evolearn.py

@@ -0,0 +1,99 @@
+#!/usr/bin/env python2.6
+# -*- coding: utf-8 -*-
+
+
+from __future__ import division
+
+
+__author__ = 'Justin S Bayer, [email protected]'
+
+
+import glob
+import itertools
+import optparse
+import os
+import sys
+
+import scipy
+from pybrain.optimization import PGPE
+
+import rnn
+
+
+
+def make_optparse():
+  parser = optparse.OptionParser()
+  parser.add_option('--datapath', dest='datapath', type='str',
+                    help='specify the directory where the data lies.')
+  parser.add_option('--hidden', dest='n_hidden', type='int',
+                    help='specify number of hiddens to use')
+  parser.add_option('--maxevals', dest='maxevals', type='int', default=100,
+                    help='specify number of maximum passes through data')
+  return parser
+
+
+def load_dataset(path):
+  sortedfilesbyglob = lambda x: sorted(glob.glob(os.path.join(path, '%s*' % x)))
+  inptfiles = sortedfilesbyglob('input')
+  targetfiles = sortedfilesbyglob('target')
+
+  data = []
+  for infn, targetfn in itertools.izip(inptfiles, targetfiles):
+    inpt = scipy.loadtxt(infn)
+    target = scipy.loadtxt(targetfn)
+    target.shape = scipy.size(target), 1
+    data.append((inpt, target))
+  return data
+
+
+def make_objective_func(net, data, errorfunc):
+  def obj(x):
+    net.parameters[:] = x
+    error = 0
+    for inpt, target in data:
+      _, output = net(inpt)
+      error += float(errorfunc(output, target))
+    return error
+  return obj
+
+
+def stats(net, data):
+  true_positives = 0
+  total = 0
+  for inpt, target in data:
+    _, output = net(inpt)
+    output = (output > 0.5).astype('float64')
+    total += target.shape[0]
+    true_positives += (output * target).sum()
+  return true_positives, total
+
+
+def main():
+  options, args = make_optparse().parse_args()
+  print "Loading data"
+  data = load_dataset(options.datapath)
+
+  print "Building network"
+  net = rnn.RecurrentNetwork(74, options.n_hidden, 1, outfunc='sig')
+  print "Number of parameters:", len(net.parameters)
+
+  objfunc = make_objective_func(net, data, rnn.SumOfSquares())
+  x0 = scipy.random.standard_normal(len(net.parameters)) * 0.1
+  optimizer = PGPE(objfunc, x0, minimize=True)
+  optimizer.maxEvaluations = options.maxevals
+
+  print "First fitness:", objfunc(x0)
+  print "Optimizing..."
+  params, fitness = optimizer.learn()
+  print "Last fitness:", fitness
+
+  true_positives, total = stats(net, data)
+  print "Total positives found: %i (%.2f)" % (true_positives, true_positives /
+                                                total)
+  
+  return 0
+
+
+if __name__ == '__main__':
+  sys.exit(main())
+