much things

Author: tariqdaouda

Mariana/candies.py

@@ -37,7 +37,7 @@ def fatal(msgSubject, msg, toRaise = ValueError, flush = True) :
 
 	subject = msgSubject
 
-	s = """\n%s:\n%s\n%s\n\n  Sorry,\n\n  Mariana\n""" %(subject, "-"*(len(subject) + 1), m)
+	s = """\n%s:\n%s\n%s\n\n  %s\nSorry,\n\n  Mariana\n""" %(subject, "-"*(len(subject) + 1), m, toRaise.message)
 
 	if MSET.SAVE_MESSAGE_LOG :
 		if not MESSAGE_LOG_FILE :
@@ -46,4 +46,4 @@ def fatal(msgSubject, msg, toRaise = ValueError, flush = True) :
 		if flush :
 			MESSAGE_LOG_FILE.flush()
 
-	raise toRaise(s)
+	raise toRaise

Mariana/doc/source/conf.py

@@ -64,9 +64,9 @@
 # built documents.
 #
 # The short X.Y version.
-version = '1.0.1'
+version = '1.0.x'
 # The full version, including alpha/beta/rc tags.
-release = '1.0.1rc'
+release = '1.0.xrc'
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.

Mariana/initializations.py

@@ -7,6 +7,7 @@
 __all__= [
 	"Initialization_ABC",
 	"HardSet",
+	"Identity",
 	"GlorotTanhInit",
 	"Uniform",
 	"UniformWeights",
@@ -44,6 +45,15 @@ def initialize(self, layer) :
 		"""The function that all Initialization_ABCs must implement"""
 		raise NotImplemented("This one should be implemented in child")
 
+class Identity(Initialization_ABC) :
+	"""Identity matrix for weights"""
+	def __init__(self, *args, **kwargs) :
+		Initialization_ABC.__init__(self, *args, **kwargs)
+		
+	def initialize(self, layer) :
+		v = numpy.identity(layer.nbOutputs, dtype = theano.config.floatX)
+		setattr( layer, "W",  theano.shared(value = v, name = "%s_%s" % (layer.name, "W") ) )
+
 class HardSet(Initialization_ABC) :
 	"""Sets the parameter to value (must have a correct shape)"""
 	def __init__(self, parameter, value, *args, **kwargs) :

Mariana/layers.py

@@ -168,7 +168,7 @@ def _setTheanoFunctions(self) :
 		propagateTest returns the testOutputs, some decorators might not be applied.
 		This is called after decorating"""
 		self.propagate = MWRAP.TheanoFunction("propagate", self, [("outputs", self.outputs)], allow_input_downcast=True)
-		self.propagateTest = MWRAP.TheanoFunction("propagateTest", self, [("testOutputs", self.testOutputs)], allow_input_downcast=True)
+		self.propagateTest = MWRAP.TheanoFunction("propagateTest", self, [("outputs", self.testOutputs)], allow_input_downcast=True)
 
 	def setCustomTheanoFunctions(self) :
 		"""This is where you should put the definitions of your custom theano functions. Theano functions
@@ -523,11 +523,11 @@ def setCustomTheanoFunctions(self) :
 		clasAcc = tt.mean( tt.eq(self.targets, clas ) )
 		predAcc = tt.mean( tt.eq(self.targets, pred ) )
 
-		self.classificationAccuracy = MWRAP.TheanoFunction("accuracy", self, [("accuracy", clasAcc)], { "targets" : self.targets }, allow_input_downcast=True)
-		self.predictionAccuracy = MWRAP.TheanoFunction("accuracy", self, [("accuracy", predAcc)], { "targets" : self.targets }, allow_input_downcast=True)
+		self.classificationAccuracy = MWRAP.TheanoFunction("classificationAccuracy", self, [("accuracy", clasAcc)], { "targets" : self.targets }, allow_input_downcast=True)
+		self.predictionAccuracy = MWRAP.TheanoFunction("predictionAccuracy", self, [("accuracy", predAcc)], { "targets" : self.targets }, allow_input_downcast=True)
 
-		self.trainAndAccuracy = MWRAP.TheanoFunction("accuracy", self, [("score", self.cost), ("accuracy", clasAcc)], { "targets" : self.targets },  updates = self.updates, allow_input_downcast=True)
-		self.testAndAccuracy = MWRAP.TheanoFunction("accuracy", self, [("score", self.testCost), ("accuracy", predAcc)], { "targets" : self.targets }, allow_input_downcast=True)
+		self.trainAndAccuracy = MWRAP.TheanoFunction("trainAndAccuracy", self, [("score", self.cost), ("accuracy", clasAcc)], { "targets" : self.targets },  updates = self.updates, allow_input_downcast=True)
+		self.testAndAccuracy = MWRAP.TheanoFunction("testAndAccuracy", self, [("score", self.testCost), ("accuracy", predAcc)], { "targets" : self.targets }, allow_input_downcast=True)
 
 class Regression(Output_ABC) :
 	"""For regressions, works great with a mean squared error cost"""

Mariana/network.py

@@ -52,10 +52,10 @@ def __repr__(self) :
 			os.append(o.name)
 
 		os = ', '.join(os)
-		return "<theano fct '%s' for output layer: '%s'>" % (self.name, os)
+		return "<theano fct '%s' for layers: '%s'>" % (self.name, os)
 
 class Network(object) :
-	"""All **theano_x** functions of the outputs are accessible through the network interface **network.x(...)**."""
+	"""All theano functions of all layers are accessible through the network interface **network.x(...)**."""
 	def __init__(self) :
 		self.inputs = OrderedDict()
 		self.layers = OrderedDict()
@@ -227,7 +227,7 @@ def help(self) :
 			os.append(repr(o))
 		os = '\n\t'.join(os)
 
-		print "Available model functions:\n" % os
+		print "Available model functions:\n%s\n" % os
 
 	def save(self, filename) :
 		import cPickle, pickle

Mariana/regularizations.py

@@ -13,7 +13,7 @@ def apply(self, layer) :
 
 		message = "%s uses %s regularization" % (layer.name, self.__class__.__name__)
 		layer.network.logLayerEvent(layer, message, hyps)
-		return self.getFormula(layer, x)
+		return self.getFormula(layer)
 
 	def getFormula(self, layer) :
 		"""Returns the expression to be added to the cost"""

Mariana/training/recorders.py

@@ -67,7 +67,13 @@ def __init__(self, mapName, outputName, functionName) :
 			self.outputName = outputName.name
 
 		self.functionName = functionName
-	
+		self.recorder = None
+
+	def _shouldISave(self, recorder) :
+		"""This is the function that is called by the recorder. If sets self.recorder to 'recorder' and then calls self.shouldISave()"""
+		self.recorder = recorder
+		return self.shouldISave(recorder)
+
 	def shouldISave(self, recorder) :
 		"""The function that defines when a save should be performed"""
 		raise NotImplemented("Should be implemented in child")
@@ -79,7 +85,7 @@ def getFilename(self, recorder) :
 	def loadLast(self):
 		"""load the last saved model"""
 		import Mariana.network as MNET
-		return MNET.loadModel(self.getFilename())
+		return MNET.loadModel(self.getFilename(self.recorder))
 
 	def __repr__(self) :
 		return "%s on %s" %(self.__class__.__name__, (self.mapName, self.outputName, self.functionName) )
@@ -198,7 +204,7 @@ def _fillLine(csvFile, score, minScore, maxScore, mapName, setLen, outputName, o
 			self.printCurrentState()
 
 		for w in self.whenToSave :
-			if w.shouldISave(self) :
+			if w._shouldISave(self) :
 				model.save(w.getFilename(self))
 
 	def printCurrentState(self) :

Mariana/training/stopcriteria.py

@@ -27,7 +27,7 @@ def __init__(self, maxEpochs) :
 		self.maxEpochs = maxEpochs
 
 	def stop(self, trainer) :
-		if trainer.store["runInfos"]["epoch"] >= self.maxEpochs :
+		if trainer.store["runInfos"]["epoch"]+1 >= self.maxEpochs :
 			return True
 		return False
 

Mariana/training/trainers.py

@@ -276,7 +276,6 @@ def _trainTest(aMap, modelFct, trainingOrder, miniBatchSize, inputLayers, output
 							layerList.append(output)
 							batchData = aMap.getBatch(i, miniBatchSize, layerList = layerList)
 							res = modelFct(output, **batchData)
-
 							if output.name not in scores :
 								scores[output.name] = {}
 
@@ -287,7 +286,6 @@ def _trainTest(aMap, modelFct, trainingOrder, miniBatchSize, inputLayers, output
 									scores[output.name][k] = [v]
 
 							layerList.pop(-1)
-
 				elif trainingOrder == DefaultTrainer.SIMULTANEOUS_TRAINING :
 					for i in xrange(0, len(aMap), miniBatchSize) :
 						batchData = aMap.getBatch(i, miniBatchSize, layerList = layerList)
@@ -383,7 +381,7 @@ def _trainTest(aMap, modelFct, trainingOrder, miniBatchSize, inputLayers, output
 			self.store["setSizes"][mapName] = self.maps[mapName].getMinFullLength()
 			self.store["scores"][mapName] = {}
 			for o in outputLayers :
-				self.store["scores"][mapName][o.name] = "NA"
+				self.store["scores"][mapName][o.name] = {}
 
 		while True :
 			for mapName in ["train", "test", "validation"] :

Mariana/wrappers.py

@@ -96,15 +96,22 @@ def dumpToHTML(self, filename) :
 		d3v.d3viz(self.theano_fct, filename)
 
 	def run(self, **kwargs) :
-		"""Run the thano function with the kwargs. Will return an OrderedDict of the outputs"""
+		"""Run the theano function with the kwargs. Will return an OrderedDict of the outputs"""
 		def _die(fctName, outputLayer, kwargs, exc) :
-			sys.stderr.write("!!=> Error in function '%s' for layer '%s':\n" % (fctName, outputLayer.name))
-			sys.stderr.write("\t!!=> the arguments were:\n %s\n" % (kwargs))
-			raise exc
+			localStr = "!!=> Error in function '%s' for layer '%s':\n%s\n" % (fctName, outputLayer.name, exc.message)
+			sys.stderr.write(localStr)
+			sys.stderr.write("Have a look at the log file: %s for details about the arguments" % MSET.SAVE_MESSAGE_LOG_FILE)
+			strArgs = []
+			for k, v in kwargs.iteritems() :
+				strArgs.append("%s, shape: %s \n----\n%s" % (k, numpy.asarray(v).shape, v))
+			MCAN.fatal(localStr, "!!=> the arguments were:\n %s\n" % ('\n'.join(strArgs)), toRaise = exc)
 
 		self.fctInputs.update(kwargs)
-		fres = iter(self.theano_fct(*self.fctInputs.values()))
-		
+		try :
+			fres = iter(self.theano_fct(*self.fctInputs.values()))
+		except Exception as e:
+			_die(self.name, self.outputLayer, kwargs, e)
+			
 		for k in self.output_expressions.iterkeys() :
 			self.results[k] = fres.next()