def createNetworkLayout(logger, preprocessor):
'''
Returns the network with the specified layout.
'''
# Create Neural Network
network = NeuralNetwork()
network.createSequentialModel()
input_shape = (preprocessor.getNetworkData()['input'].shape[1],
preprocessor.getNetworkData()['input'].shape[2])
vokab_length = len(preprocessor.getLabelEncoder().classes_)
# Add Layers
# units = how many nodes a layer should have
# input_shape = shape of the data it will be training
network.add(LSTM(units=256, input_shape=input_shape,
return_sequences=True))
# rate = fraction of input units that should be dropped during training
network.add(Dropout(rate=0.3))
network.add(LSTM(units=512, return_sequences=True))
network.add(Dropout(rate=0.3))
network.add(LSTM(units=256))
network.add(Dense(units=256))
network.add(Dropout(rate=0.3))
# units of last layer should have same amount of nodes as the number of different outputs that our system has
# -> assures that the output of the network will map to our classes
network.add(Dense(units=vokab_length))
network.add(Activation('softmax'))
logger.info("Compiling model...")
network.compile(_loss='categorical_crossentropy',
_optimizer='rmsprop',
_metrics=['acc'])
logger.info("Finished compiling.")
logger.info("Model Layers: \n[]".format(network._model.summary()))
return network
def createNetworkLayout(logger,
preprocessor,
layout,
loss,
optimizer,
activation,
metrics,
weightsPath=None,
dropout=0.3,
callbacks=[]):
'''
Creates the network layout.
Will validate the weightsPath so you dont have to take care of that.
Returns the network with the specified layout.
'''
# check for correctness and create folder if missing
if not weightsPath is None:
weightsPath = validateFolderPath(weightsPath, logger)
# Create Neural Network
network = NeuralNetwork()
network.createSequentialModel()
input_shape = (preprocessor.getNetworkData()['input'].shape[1],
preprocessor.getNetworkData()['input'].shape[2])
vokab_length = len(preprocessor.getLabelEncoder().classes_)
# Add Layers
# units = how many nodes a layer should have
# input_shape = shape of the data it will be training
layout = layout
if layout == 'default':
network = defaultLayout(network, input_shape, dropout)
elif layout == 'multi':
network = multiLSTMLayout(network, input_shape, dropout)
elif layout == 'bidirectional':
network = bidirectionalLayout(network, input_shape, dropout)
elif layout == 'multibidirectional':
network = multibidirectionalLayout(network, input_shape, dropout)
#elif layout == 'attention':
# network = attentionLayout(network, input_shape)
# units of last layer should have same amount of nodes as the number of different outputs that our system has
# last layers are the same for every layout
# -> assures that the output of the network will map to our classes
network.add(Dense(units=vokab_length))
network.add(Activation(activation))
# compile network
logger.info("Compiling model...")
network.compile(_loss=loss,
_path=weightsPath,
_optimizer=optimizer,
_metrics=metrics,
_callbacks=callbacks)
logger.info("Finished compiling.")
#logger.info("Model Layers: \n[]".format(network._model.summary()))
return network