network model functions now take the output layer as first argument.

dataset mapper will have to be rewritten, mad more simple

Author: tariqdaouda

.gitignore

@@ -52,3 +52,9 @@ docs/_build/
 
 # PyBuilder
 target/
+
+# my stuff
+test/
+trash/
+Mariana/test/
+Mariana/trash

Mariana/examples/mnist_mlp.py

@@ -33,14 +33,12 @@ def load_mnist() :
 
 if __name__ == "__main__" :
 
-	miniBatchSize = 20
-
 	#Let's define the network
 	ls = MS.DefaultScenario(lr = 0.01, momentum = 0)
 	cost = MC.NegativeLogLikelihood()
 
-	i = ML.Input(28*28, 'inp')
-	h = ML.Hidden(500, activation = MA.tanh, decorators = [MD.GlorotTanhInit()], regularizations = [ MR.L1(0), MR.L2(0.0001) ] )
+	i = ML.Input(28*28, name = 'inp')
+	h = ML.Hidden(500, activation = MA.tanh, decorators = [MD.GlorotTanhInit()], regularizations = [ MR.L1(0), MR.L2(0.0001) ], name = "hid" )
 	o = ML.SoftmaxClassifier(10, learningScenario = ls, costObject = cost, name = "out", regularizations = [ MR.L1(0), MR.L2(0.0001) ] )
 
 	mlp = i > h > o
@@ -69,7 +67,7 @@ def load_mnist() :
 		testMaps = testMaps,
 		validationMaps = validationMaps,
 		stopCriteria = [earlyStop, epochWall],
-		trainMiniBatchSize = miniBatchSize
+		trainMiniBatchSize = 20
 	)
 
-	trainer.start("MLP", mlp)
+	trainer.start("MLP", mlp, shuffleMinibatches = False)

Mariana/layers.py

@@ -18,7 +18,7 @@ class Layer_ABC(object) :
 
 	__metaclass__ = ABCMeta
 
-	def __init__(self, nbOutputs, saveOutputs = True, decorators = [], name = None) :
+	def __init__(self, nbOutputs, saveOutputs = False, decorators = [], name = None) :
 
 		if name is not None :
 			self.name = name
@@ -315,7 +315,7 @@ def _setOutputs(self) :
 				updates.extend(self.learningScenario.getUpdates(l, cost))
 
 		for l in self.network.layers.itervalues() :
-			if l.last_outputs is not l.outputs is not None :
+			if ( l.last_outputs is not None ) and ( l.outputs is not None ) :
 				updates.append( (l.last_outputs, l.outputs ) )
 
 		self.train = TheanoFunction("train", self, [cost, self.outputs], { "target" : self.target }, updates = updates)

Mariana/network.py

@@ -11,25 +11,25 @@ def __init__(self, name):
 		self.name = name
 		self.outputFcts = {}
 
-	def printGraph(self, outputName) :
+	def printGraph(self, outputLayer) :
 		"""Print the theano graph of the function associated with a given output"""
-		self.outputFcts[outputName].printGraph()
+		self.outputFcts[outputLayer.name].printGraph()
 
-	def addOutput(self, output, fct) :
-		self.outputFcts[output.name] = fct
+	def addOutput(self, outputLayer, fct) :
+		self.outputFcts[outputLayer.name] = fct
 
-	def map(self, outputLayerName, **kwargs) :
-		return self.outputFcts[outputLayerName](**kwargs)
+	def callTheanoFct(self, outputLayer, **kwargs) :
+		return self.outputFcts[outputLayer.name](**kwargs)
 
-	def __call__(self, outputLayerName, **kwargs) :
-		return self.map(outputLayerName, **kwargs)
+	def __call__(self, outputLayer, **kwargs) :
+		return self.callTheanoFct(outputLayer, **kwargs)
 
 	def __repr__(self) :
 		os = []
 		for o, v in self.outputFcts.iteritems() :
 			os.append(o)
 		os = ', '.join(os)
-		return "<'%s' for outputs: %s>" % (self.name, os)
+		return "<theano fct '%s' for output layer: '%s'>" % (self.name, os)
 
 class Network(object) :
 	"""All theano_x functions of the outputs are accessible through the network interface network.x().

Mariana/tests/tests.py

@@ -1,7 +1,6 @@
 import unittest
 
-from Mariana.layers import *
-from Mariana.rules import *
+import Mariana.layers as ML
 import Mariana.decorators as dec
 import Mariana.costs as MC
 import Mariana.regularizations as MR
@@ -30,27 +29,29 @@ def trainMLP_xor(self) :
 		ls = MS.DefaultScenario(lr = 0.1, momentum = 0)
 		cost = MC.NegativeLogLikelihood()
 
-		i = Input(2, 'inp')
-		h = Hidden(4, activation = MA.tanh, decorators = [dec.GlorotTanhInit()], regularizations = [MR.L1(0), MR.L2(0)])
-		o = SoftmaxClassifier(2, learningScenario = ls, costObject = cost, name = "out")
+		i = ML.Input(2, 'inp')
+		h = ML.Hidden(4, activation = MA.tanh, decorators = [dec.GlorotTanhInit()], regularizations = [MR.L1(0), MR.L2(0)])
+		o = ML.SoftmaxClassifier(2, learningScenario = ls, costObject = cost, name = "out")
 
 		mlp = i > h > o
 
 		self.xor_ins = N.array(self.xor_ins)
 		self.xor_outs = N.array(self.xor_outs)
 		for i in xrange(1000) :
 			ii = i%len(self.xor_ins)
-			mlp.train("out", inp = [ self.xor_ins[ ii ] ], target = [ self.xor_outs[ ii ] ] )
+			mlp.train(o, inp = [ self.xor_ins[ ii ] ], target = [ self.xor_outs[ ii ] ] )
 
 		return mlp
 
 	# @unittest.skip("skipping")
 	def test_xor(self) :
 		mlp = self.trainMLP_xor()
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[0] ] )[0], 0 )
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[1] ] )[0], 1 )
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[2] ] )[0], 1 )
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[3] ] )[0], 0 )
+		o = mlp.outputs.values()[0]
+		
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[0] ] )[0], 0 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[1] ] )[0], 1 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[2] ] )[0], 1 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[3] ] )[0], 0 )
 
 	# @unittest.skip("skipping")
 	def test_save_load(self) :
@@ -60,10 +61,12 @@ def test_save_load(self) :
 		mlp.save("test_save")
 		mlp2 = cPickle.load(open('test_save.mariana.pkl'))
 
-		self.assertEqual(mlp2.classify( "out", inp = [ self.xor_ins[0] ] )[0], 0 )
-		self.assertEqual(mlp2.classify( "out", inp = [ self.xor_ins[1] ] )[0], 1 )
-		self.assertEqual(mlp2.classify( "out", inp = [ self.xor_ins[2] ] )[0], 1 )
-		self.assertEqual(mlp2.classify( "out", inp = [ self.xor_ins[3] ] )[0], 0 )
+
+		o = mlp.outputs.values()[0]
+		self.assertEqual(mlp2.classify( o, inp = [ self.xor_ins[0] ] )[0], 0 )
+		self.assertEqual(mlp2.classify( o, inp = [ self.xor_ins[1] ] )[0], 1 )
+		self.assertEqual(mlp2.classify( o, inp = [ self.xor_ins[2] ] )[0], 1 )
+		self.assertEqual(mlp2.classify( o, inp = [ self.xor_ins[3] ] )[0], 0 )
 
 		os.remove('test_save.mariana.pkl')
 
@@ -72,11 +75,11 @@ def test_composite(self) :
 		ls = MS.DefaultScenario(lr = 0.1, momentum = 0)
 		cost = MC.NegativeLogLikelihood()
 
-		inp = Input(2, 'inp')
-		h1 = Hidden(2, activation = MA.tanh, name = "h1")
-		h2 = Hidden(2, activation = MA.tanh, name = "h2")
-		o = SoftmaxClassifier(2, learningScenario = ls, costObject = cost, name = "out")
-		c = Composite(name = "Comp")
+		inp = ML.Input(2, 'inp')
+		h1 = ML.Hidden(2, activation = MA.tanh, name = "h1")
+		h2 = ML.Hidden(2, activation = MA.tanh, name = "h2")
+		o = ML.SoftmaxClassifier(2, learningScenario = ls, costObject = cost, name = "out")
+		c = ML.Composite(name = "Comp")
 
 		inp > h1 > c
 		inp > h2 > c
@@ -86,12 +89,12 @@ def test_composite(self) :
 		self.xor_outs = N.array(self.xor_outs)
 		for i in xrange(1000) :
 			ii = i%len(self.xor_ins)
-			mlp.train("out", inp = [ self.xor_ins[ ii ] ], target = [ self.xor_outs[ ii ] ])
+			mlp.train(o, inp = [ self.xor_ins[ ii ] ], target = [ self.xor_outs[ ii ] ])
 
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[0] ] )[0], 0 )
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[1] ] )[0], 1 )
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[2] ] )[0], 1 )
-		self.assertEqual(mlp.classify( "out", inp = [ self.xor_ins[3] ] )[0], 0 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[0] ] )[0], 0 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[1] ] )[0], 1 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[2] ] )[0], 1 )
+		self.assertEqual(mlp.classify( o, inp = [ self.xor_ins[3] ] )[0], 0 )
 
 if __name__ == '__main__' :
 	unittest.main()

Mariana/training/datasetmaps.py

@@ -2,9 +2,12 @@
 import numpy, random, time
 
 class ListSet(object) :
-	def __init__(self, values) :
+	def __init__(self, values, name = None) :
 		self.values = values
-		self.name = time.clock() + random.randint(0, 100)
+		if name is not None :
+			self.name = name
+		else :
+			self.name = time.clock() + random.randint(0, 100)
 
 	def getAll(self) :
 		return self.values
@@ -149,6 +152,7 @@ def __init__(self):
 
 		self.minLen = 0
 		self.runIds = []
+		self.mustInit = True
 
 	def mapInput(self, lst, layer) :
 		if layer.name in self.layerNames :
@@ -210,6 +214,9 @@ def syncLayers(self, refLayer, layer) :
 		self.layerNames.add(layer.name)
 
 	def shuffle(self) :
+		if self.mustInit :
+			self._init()
+
 		for s in self.sets.itervalues() :
 			if s.__class__ is ClassSuperset : 
 				s.setMinLen(self.minLen)
@@ -219,29 +226,35 @@ def shuffle(self) :
 			self.runIds = range(self.minLen)
 		random.shuffle(self.runIds)
 
+	def _init(self) :
+		self.runIds = range(self.minLen)
+		self.mustInit = False
+
  	def getBatches(self, i, size) :
 		"""Returns a random set of examples for each class, all classes have an equal chance of apperance
 		regardless of their number of elements. If you want  the limit to be length of the whole set
  		instead of a mini batch you can set size to "all".
  		"""
-
+		if self.mustInit :
+			self._init()
+			
  		inps = {}
 		outs = {}
-		for ii in self.runIds[i: i+size] :
-			for k, v in self.sets.iteritems() :
-				elmt = v[ii: ii+size]
-				if v.__class__ is ClassSuperset :
+		ii = self.runIds[i]
+		for k, v in self.sets.iteritems() :
+			elmt = v[ii: ii+size]
+			if v.__class__ is ClassSuperset :
+				l = self.inputSets[k]
+				inps[l.name] = elmt[0]
+				l = self.outputSets[k]
+				outs[l.name] = elmt[1]
+			else :
+				try :
 					l = self.inputSets[k]
-					inps[l.name] = elmt[0]
+					inps[l.name] = elmt
+				except :
 					l = self.outputSets[k]
-					outs[l.name] = elmt[1]
-				else :
-					try :
-						l = self.inputSets[k]
-						inps[l.name] = elmt
-					except :
-						l = self.outputSets[k]
-						outs[l.name] = elmt
+					outs[l.name] = elmt
 
 		for k, layers in self.syncedLayers.iteritems() :
 			for l in layers :
@@ -257,12 +270,15 @@ def getBatches(self, i, size) :
 					except KeyError :
 						outs[l.name] = outs[k]
 				else :
-					raise ValueError("Synced layer ''%s is neither an input nor an output" % l.name)
+					raise ValueError("Synced layer ''%s is neither an input nor an output layer" % l.name)
 
 		return (inps, outs)
 
 	def getAll(self) :
 		"""Returns the whole batch"""
+		if self.mustInit :
+			self._init()
+
 		inps = {}
 		outs = {}
 
@@ -294,12 +310,15 @@ def getAll(self) :
 					except KeyError :
 						outs[l.name] = outs[k]
 				else :
-					raise ValueError("Synced layer ''%s is neither an input nor an output" % l.name)
+					raise ValueError("Synced layer ''%s is neither an input nor an output layer" % l.name)
 
 		return (inps, outs)
 
 	def getOutputNames(self) :
  		return self.outputLayerNames
 
+ 	def __repr__(self) :
+ 		return "<DatasetMapper len: %s, sets: %s, inputs: %s, outputs: %s>" % (self.minLen, len(self.sets), len(self.inputSets), len(self.outputSets))
+
 	def __len__(self) :
 		return self.minLen

Mariana/training/recoders.py

@@ -1,4 +1,4 @@
-import sys
+import sys, os
 from pyGeno.tools.parsers.CSVTools import CSVFile
 
 class Recorder_ABC(object) :
@@ -21,7 +21,7 @@ def __init__(self, filename, verbose = True):
 
  		self.bestScores = {}
 		self.currentScores = {}
-		
+
 		self.csvLegend = None
 		self.csvFile = None
 
@@ -43,7 +43,8 @@ def _fillLine(csvFile, score, bestScore, setName, setLen, outputName, **csvValue
 		self.length += 1
 		if self.csvLegend is None :
 			self.csvLegend = store["hyperParameters"].keys()
-			self.csvLegend.extend( ["score", "best_score", "set", "output"] )
+			self.csvLegend.extend(store["runInfos"].keys())
+			self.csvLegend.extend( ["score", "best_score", "best_score_commit", "set", "output"] )
 
 			self.csvFile = CSVFile(legend = self.csvLegend)
 			self.csvFile.streamToFile( self.filename, writeRate = 1 )
@@ -58,14 +59,16 @@ def _fillLine(csvFile, score, bestScore, setName, setLen, outputName, **csvValue
 					self.bestScores[theSet][outputName] = (score, self.length)
 					model.save("best-%s-%s" % (theSet, self.filename))
 
+				muchData = store["hyperParameters"]
+				muchData.update(store["runInfos"]) 
 				_fillLine(
 					self.csvFile,
 					self.currentScores[theSet][outputName],
 					self.bestScores[theSet][outputName],
 					theSet,
 					store["setSizes"][theSet],
 					outputName,
-					**store["hyperParameters"]
+					**muchData
 				)
 
 
@@ -75,7 +78,7 @@ def _fillLine(csvFile, score, bestScore, setName, setLen, outputName, **csvValue
 	def printCurrentState(self) :
 		"""prints the current state stored in the recorder"""
 		if self.length > 0 :
-			print "\n==>rec: ggplot2, commit %s:" % self.length
+			print "\n==>rec: ggplot2, commit %s, pid: %s:" % (self.length, os.getpid())
 			for setName, scores in self.bestScores.iteritems() :
 				print "  |-%s set" % setName
 				for outputName in scores :