def __setattr__(self, k, v):
if k == "name" and hasattr(self, k):
if len(self.network.layers) > 1:
raise ValueError(
"You can't change the name of a connected layer")
else:
object.__setattr__(self, k, v)
self.network = MNET.Network()
self.network._addLayer(self)
try:
deco = self._decorating
except AttributeError:
object.__setattr__(self, k, v)
return
if deco:
var = getattr(self, k)
try:
var.set_value(numpy.asarray(v, dtype=theano.config.floatX),
borrow=True)
return
except AttributeError:
pass
object.__setattr__(self, k, v)
def __init__(self, size, name=None, **kwargs):
Layer_ABC.__init__(self, size, name=name, **kwargs)
self.kwargs = kwargs
self.type = TYPE_INPUT_LAYER
self.nbInputs = size
self.network = MNET.Network()
self.network.addInput(self)
self.inputs = tt.matrix(name=self.name)
def _resetNetwork(self, fullReset=True, newNetwork=None):
if fullReset:
self._initStatus = 0
if newNetwork is None:
self.network = MNET.Network()
self.network._addLayer(self)
else:
self.network = newNetwork
def __init__(self,
size,
layerTypes,
activation=MA.Pass(),
regularizations=[],
initializations=[],
learningScenario=None,
decorators=[],
name=None):
self.isLayer = True
#a unique tag associated to the layer
self.appelido = numpy.random.random()
if name is not None:
self.name = name
else:
self.name = "%s_%s" % (self.__class__.__name__, self.appelido)
self.types = layerTypes
self.nbInputs = None
self.inputs = None
self.nbOutputs = size
self.outputs = None # this is a symbolic var
self.testOutputs = None # this is a symbolic var
self.preactivation_outputs = None
self.preactivation_testOutputs = None
self.activation = activation
self.regularizationObjects = regularizations
self.regularizations = []
self.decorators = decorators
self.initializations = initializations
self.learningScenario = learningScenario
self.network = MNET.Network()
self.network._addLayer(self)
self._inputRegistrations = set()
self._mustInit = True
self._mustReset = True
self._decorating = False
self.parameters = {}
def __init__(self, size, nbDimentions, dictSize, name=None, **kwargs):
"""
:param size int: the size of the input vector (if your input is a sentence this should be the number of words in it).
:param nbDimentions int: the number of dimentions in wich to encode each word.
:param dictSize int: the total number of words.
"""
Layer_ABC.__init__(self, size, nbDimentions, name=name, **kwargs)
self.network = MNET.Network()
self.network.addInput(self)
self.type = TYPE_INPUT_LAYER
self.dictSize = dictSize
self.nbDimentions = nbDimentions
self.nbInputs = size
self.nbOutputs = self.nbDimentions * self.nbInputs
initEmb = numpy.asarray(numpy.random.random(
(self.dictSize, self.nbDimentions)),
dtype=theano.config.floatX)
self.embeddings = theano.shared(initEmb)
self.inputs = tt.imatrix()
self.test_inputs = tt.imatrix()