Added code for benchmarking pybrain, arac and theano RNN implementations.

Author: bayerj

util/speedtest.py

@@ -0,0 +1,100 @@
+#!/usr/bin/env python2.6
+# -*- coding: utf-8 -*-
+
+
+__author__ = 'Justin S Bayer, [email protected]'
+
+
+import optparse
+import sys
+import time
+
+import scipy
+
+from rnn import RecurrentNetwork, LstmNetwork
+from pybrain.tools.shortcuts import buildNetwork
+from pybrain.structure import LSTMLayer, TanhLayer
+
+
+def make_optparse():
+  parser = optparse.OptionParser()
+  parser.add_option('--lstm', dest='lstm', action='store_true')
+  return parser
+
+
+def theano(num_inpt, num_hidden, num_output, inpts, lstm=False):
+  klass = LstmNetwork if lstm else RecurrentNetwork
+  rnn = klass(num_inpt, num_hidden, num_output)
+
+  start = time.time()
+  for inpt in inpts:
+    rnn(inpt)
+  return time.time() - start
+
+
+def pybrain(num_inpt, num_hidden, num_output, inpts, lstm=False):
+  net = buildNetwork(num_inpt, num_hidden, num_output, recurrent=True,
+                     hiddenclass=LSTMLayer if lstm else TanhLayer,
+                     fast=False)
+  start = time.time()
+  for seq in inpts:
+    net.reset()
+    for inpt in seq:
+      net.activate(inpt)
+  return time.time() - start
+
+
+def pybrainarac(num_inpt, num_hidden, num_output, inpts, lstm=False):
+  net = buildNetwork(num_inpt, num_hidden, num_output, recurrent=True,
+                     hiddenclass=LSTMLayer if lstm else TanhLayer,
+                     fast=True)
+  start = time.time()
+  for seq in inpts:
+    net.reset()
+    for inpt in seq:
+      net.activate(inpt)
+  return time.time() - start
+
+
+def main():
+  options, args = make_optparse().parse_args()
+  num_inpt = int(args[0])
+  num_hidden = int(args[1])
+  num_output = int(args[2])
+
+  print "Network stats"
+  print "-" * 20
+  print "Number of inputs: %i" % num_inpt
+  print "Number of hidden: %i" % num_hidden
+  print "Number of outputs: %i" % num_output
+
+  inpts = scipy.random.random((500, 100, num_inpt))
+
+  print
+  print "Durations"
+  print "-" * 20
+  pybrain_dur = pybrain(num_inpt, num_hidden, num_output, inpts,
+                        lstm=options.lstm)
+  print "Pybrain: %.2f" % pybrain_dur
+
+  pybrainarac_dur = pybrainarac(num_inpt, num_hidden, num_output, inpts,
+                                lstm=options.lstm)
+  print "Pybrain \w arac: %.2f" % pybrainarac_dur
+
+  theano_dur = theano(num_inpt, num_hidden, num_output, inpts, lstm=options.lstm)
+  print "Theano: %.2f" % theano_dur
+
+  print
+  print "Ratios"
+  print "-" * 20
+
+  print "Theano / PyBrain: %.2f" % (theano_dur / pybrain_dur)
+  print "Theano / PyBrain+arac: %.2f" % (theano_dur / pybrainarac_dur)
+  print "PyBrain+arac / PyBrain: %.2f" % (pybrainarac_dur / pybrain_dur)
+
+  return 0
+
+
+if __name__ == '__main__':
+  sys.exit(main())
+