def __setUpBrain(self, genome):
"""
Set up PyBrain's neural network
Args:
genome (G1DList): PyEvolve's individual container
"""
self.network = FeedForwardNetwork()
inLayer = TanhLayer(14)
hiddenLayer = TanhLayer(12)
hiddenLayer2 = TanhLayer(6)
outLayer = TanhLayer(2)
self.network.addInputModule(inLayer)
self.network.addModule(hiddenLayer)
self.network.addModule(hiddenLayer2)
self.network.addOutputModule(outLayer)
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_hidden2 = FullConnection(hiddenLayer, hiddenLayer2)
hidden2_to_out = FullConnection(hiddenLayer2, outLayer)
self.network.addConnection(in_to_hidden)
self.network.addConnection(hidden_to_hidden2)
self.network.addConnection(hidden2_to_out)
self.network.sortModules()
new_params = numpy.array(genome.genomeList)
self.network._setParameters(new_params)
def getNetwork(trndata):
n = RecurrentNetwork()
n.addInputModule(LinearLayer(trndata.indim, name='in'))
n.addModule(SigmoidLayer(100, name='hidden'))
n.addOutputModule(LinearLayer(trndata.outdim, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(
FullConnection(n['hidden'], n['hidden'], name='c3'))
n.sortModules()
# fnn = buildNetwork( trndata.indim, 5, trndata.outdim, outclass=SoftmaxLayer )
trainer = BackpropTrainer(n,
dataset=trndata,
momentum=0.1,
verbose=True,
weightdecay=0.01)
# TODO: return network and trainer here. Make another function for training
# for i in range(20):
# trainer.trainEpochs(1)
# trainer.trainUntilConvergence(maxEpochs=100)
# trnresult = percentError( trainer.testOnClassData(),trndata['class'] )
# tstresult = percentError( trainer.testOnClassData(dataset=tstdata ), tstdata['class'] )
# print "epoch: %4d" % trainer.totalepochs, \
# " train error: %5.2f%%" % trnresult
# out = fnn.activateOnDataset(tstdata)
# out = out.argmax(axis=1) # the highest output activation gives the class
return (n, trainer)
def setupNetwork(numHiddenNodes, numHiddenLayers, numFeatures, numSpeakers):
nn = FeedForwardNetwork()
inputLayer = LinearLayer(numFeatures)
nn.addInputModule(inputLayer)
hiddenLayers = []
for x in range(numHiddenLayers):
hiddenLayer = TanhLayer(numHiddenNodes)
nn.addModule(hiddenLayer)
hiddenLayers.append(hiddenLayer)
outputLayer = SoftmaxLayer(numSpeakers)
nn.addOutputModule(outputLayer)
inputConnection = FullConnection(inputLayer, hiddenLayers[0])
nn.addConnection(inputConnection)
for x in range(numHiddenLayers - 1):
connect = FullConnection(hiddenLayers[x], hiddenLayers[x - 1])
nn.addConnection(connect)
outputConnection = FullConnection(hiddenLayers[numHiddenLayers - 1],
outputLayer)
nn.addConnection(outputConnection)
nn.sortModules()
return nn
def crearRN():
#Se crea la red neuronal
n = FeedForwardNetwork()
#Se declaran las laminas de entrada, las laminas escondidas y las de salida de la red neuronal
inLayer = LinearLayer(4096)
hiddenLayer = SigmoidLayer(3)
outLayer = LinearLayer(1)
#Se agregan los layers a la red neuronal
n.addInputModule(inLayer)
n.addModule(hiddenLayer)
n.addOutputModule(outLayer)
#Se declaran las conexiones de los nodos
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer, outLayer)
#Se establecen las conexiones en los layers de la red neuronal
n.addConnection(in_to_hidden)
n.addConnection(hidden_to_out)
#Red neuronal lista para usar
n.sortModules()
return n
def __init__(self, genes=None):
self.net = FeedForwardNetwork()
self.inLayer = TanhLayer(16)
self.hiddenLayer = TanhLayer(20)
self.hiddenLayer2 = TanhLayer(20)
self.outLayer = SoftmaxLayer(4)
self.net.addInputModule(self.inLayer)
self.net.addModule(self.hiddenLayer)
self.net.addModule(self.hiddenLayer2)
self.net.addOutputModule(self.outLayer)
self.in_to_hidden = FullConnection(self.inLayer, self.hiddenLayer)
self.hidden1_to_hidden2 = FullConnection(self.hiddenLayer, self.hiddenLayer2)
self.hidden2_to_out = FullConnection(self.hiddenLayer2, self.outLayer)
self.net.addConnection(self.in_to_hidden)
self.net.addConnection(self.hidden1_to_hidden2)
self.net.addConnection(self.hidden2_to_out)
self.net.sortModules()
# Set the params to the provided params
if genes is not None:
self.net._setParameters(genes)
def initalize_nn():
global in_to_hidden
global hidden_to_hidden2
global hidden_to_out
# Old code (regression)
n = FeedForwardNetwork()
# n = buildNetwork( 2, 3, data.outdim, outclass=SoftmaxLayer )
inLayer = LinearLayer(2)
hiddenLayer = SigmoidLayer(3)
hiddenLayer2 = SigmoidLayer(3)
outLayer = LinearLayer(1)
n.addInputModule(inLayer)
n.addModule(hiddenLayer)
n.addModule(hiddenLayer2)
n.addOutputModule(outLayer)
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_hidden2 = FullConnection(hiddenLayer, hiddenLayer2)
hidden_to_out = FullConnection(hiddenLayer2, outLayer)
n.addConnection(in_to_hidden)
n.addConnection(hidden_to_hidden2)
n.addConnection(hidden_to_out)
n.sortModules()
return n
def build_deep_network(linear_dimensions):
neural_net = FeedForwardNetwork()
inLayer = LinearLayer(linear_dimensions)
hiddenLayer_1 = SigmoidLayer(100)
hiddenLayer_2 = SigmoidLayer(100)
hiddenLayer_3 = SigmoidLayer(50)
outLayer = LinearLayer(1)
neural_net.addInputModule(inLayer)
neural_net.addModule(hiddenLayer_1)
neural_net.addModule(hiddenLayer_2)
neural_net.addModule(hiddenLayer_3)
neural_net.addOutputModule(outLayer)
in_to_hidden_1 = FullConnection(inLayer, hiddenLayer_1)
hidden_1_to_hidden_2 = FullConnection(hiddenLayer_1, hiddenLayer_2)
hidden_2_to_hidden_3 = FullConnection(hiddenLayer_2, hiddenLayer_3)
hidden_3_to_output = FullConnection(hiddenLayer_3, outLayer)
neural_net.addConnection(in_to_hidden_1)
neural_net.addConnection(hidden_1_to_hidden_2)
neural_net.addConnection(hidden_2_to_hidden_3)
neural_net.addConnection(hidden_3_to_output)
neural_net.sortModules()
return neural_net
def create_network():
# Create the network itself
network = FeedForwardNetwork()
# Create layers
NUMBER_OF_INPUT_BYTES = 1600 # because at input we have picture 40x40 size
NUMBER_OF_HIDDEN_LAYERS = 10 # number of hidden layers
NUMBER_OF_OUTPUT_CLASSES = 8 # because in output we have 8 classes
inLayer = LinearLayer(NUMBER_OF_INPUT_BYTES)
hiddenLayer = SigmoidLayer(NUMBER_OF_HIDDEN_LAYERS)
outLayer = LinearLayer(NUMBER_OF_OUTPUT_CLASSES)
# Create connections between layers
# We create FullConnection - each neuron of one layer is connected to each neuron of other layer
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer, outLayer)
# Add layers to our network
network.addInputModule(inLayer)
network.addModule(hiddenLayer)
network.addOutputModule(outLayer)
# Add connections to network
network.addConnection(in_to_hidden)
network.addConnection(hidden_to_out)
# Sort modules to make multilayer perceptron usable
network.sortModules()
# prepare array to activate network
d_letter_array = read_array("d")
# activate network
network.activate(d_letter_array)
return network
def buildNN2HiddenLayer(trnData, netNo):
from pybrain.structure import FeedForwardNetwork, RecurrentNetwork
from pybrain.structure import LinearLayer, SigmoidLayer, TanhLayer, SoftmaxLayer
from pybrain.structure import FullConnection
n = FeedForwardNetwork()
inLayer = LinearLayer(trnData.indim) # Define Layer Types
if netNo == 1 or netNo == 3:
hiddenLayer0 = TanhLayer(hiddenLayer0neurons) # Tanh
hiddenLayer1 = SigmoidLayer(hiddenLayer1neurons) # Sigmoid
elif netNo == 2:
hiddenLayer0 = TanhLayer(hiddenLayer1neurons) # Tanh
hiddenLayer1 = SigmoidLayer(hiddenLayer0neurons) # Sigmoid
outLayer = SoftmaxLayer(trnData.outdim) # SoftmaxLayer
n.addInputModule(inLayer)
n.addModule(hiddenLayer0)
n.addModule(hiddenLayer1)
n.addOutputModule(outLayer)
in_to_hidden0 = FullConnection(inLayer, hiddenLayer0) # Define connections
hidden0_to_hidden1 = FullConnection(hiddenLayer0, hiddenLayer1)
hidden1_to_out = FullConnection(hiddenLayer1, outLayer)
n.addConnection(in_to_hidden0)
n.addConnection(hidden0_to_hidden1)
n.addConnection(hidden1_to_out)
n.sortModules()
return n
def train_net(data_set, n, epochs=1):
num_inputs = len(data_set[0][0][n])
ds = SupervisedDataSet(num_inputs, 2)
for i in range(len(data_set)):
try:
ds.appendLinked(data_set[i][0][n],
(data_set[i][1], data_set[i][2]))
except:
continue
print str(len(ds)) + ' points successfully aquired'
net = FeedForwardNetwork()
net.addInputModule(LinearLayer(num_inputs, name='input'))
net.addInputModule(BiasUnit(name='bias'))
net.addOutputModule(LinearLayer(2, name='output'))
net.addModule(SigmoidLayer(int((num_inputs + 2) / 2.), name='sigmoid'))
net.addModule(TanhLayer(10, name='tanh'))
net.addConnection(FullConnection(net['bias'], net['sigmoid']))
net.addConnection(FullConnection(net['bias'], net['tanh']))
net.addConnection(FullConnection(net['input'], net['sigmoid']))
net.addConnection(FullConnection(net['sigmoid'], net['tanh']))
net.addConnection(FullConnection(net['tanh'], net['output']))
net.sortModules()
trainer = BackpropTrainer(net,
learningrate=0.01,
momentum=0.1,
verbose=True)
trainer.trainOnDataset(ds)
trainer.trainEpochs(epochs)
return net
def _init_net(params_len, output_layer_num, hidden_size):
# init and train
net = FeedForwardNetwork()
""" Next, we're constructing the input, hidden and output layers. """
inLayer = LinearLayer(params_len)
hiddenLayer = SigmoidLayer(hidden_size)
hiddenLayer1 = SigmoidLayer(hidden_size)
outLayer = LinearLayer(output_layer_num)
""" (Note that we could also have used a hidden layer of type TanhLayer, LinearLayer, etc.)
Let's add them to the network: """
net.addInputModule(inLayer)
net.addModule(hiddenLayer)
net.addModule(hiddenLayer1)
net.addOutputModule(outLayer)
""" We still need to explicitly determine how they should be connected. For this we use the most
common connection type, which produces a full connectivity between two layers (or Modules, in general):
the 'FullConnection'. """
in2hidden = FullConnection(inLayer, hiddenLayer)
hidden2hidden = FullConnection(hiddenLayer, hiddenLayer1)
hidden2out = FullConnection(hiddenLayer1, outLayer)
net.addConnection(in2hidden)
net.addConnection(hidden2hidden)
net.addConnection(hidden2out)
""" All the elements are in place now, so we can do the final step that makes our MLP usable,
which is to call the 'sortModules()' method. """
net.sortModules()
# net = buildNetwork( params_len, hidden_size, 601, bias = True )
return net
def create(number_of_hidden_layers, activation_function, input_length,
output_length, network_file, classify):
n = FeedForwardNetwork()
in_layer = LinearLayer(input_length)
n.addInputModule(in_layer)
layer_to_connect_to = in_layer
for x in range(0, number_of_hidden_layers):
if activation_function == 'sigmoid':
hidden_layer = SigmoidLayer(input_length)
else:
hidden_layer = TanhLayer(input_length)
n.addModule(hidden_layer)
hidden_layer_connection = FullConnection(layer_to_connect_to,
hidden_layer)
n.addConnection(hidden_layer_connection)
layer_to_connect_to = hidden_layer
if classify:
out_layer = SoftmaxLayer(output_length)
else:
out_layer = LinearLayer(output_length)
n.addOutputModule(out_layer)
hidden_to_out = FullConnection(layer_to_connect_to, out_layer)
n.addConnection(hidden_to_out)
n.sortModules()
save_network(n, network_file)
def __init__(self):
self.Q = FeedForwardNetwork()
# La funcion de valor se representa con una red neuronal
# Input: S = (Angulo, Velocidad angular, Posicion), A = accion
# Output: Valor
# 2 capas ocultas de 5 neuronas cada una
# Funcion de activacion sigmoidea
inLayer = SigmoidLayer(4, name="Input Layer")
hiddenLayer1 = SigmoidLayer(5, name="Hidden Layer 1")
hiddenLayer2 = SigmoidLayer(5, name="Hidden Layer 2")
outLayer = SigmoidLayer(1, name="Output Layer")
self.Q.addInputModule(inLayer)
self.Q.addModule(hiddenLayer1)
self.Q.addModule(hiddenLayer2)
self.Q.addOutputModule(outLayer)
connInToHidden1 = FullConnection(inLayer, hiddenLayer1)
connHidden1ToHidden2 = FullConnection(hiddenLayer1, hiddenLayer2)
connHidden2ToOut = FullConnection(hiddenLayer2, outLayer)
self.Q.addConnection(connInToHidden1)
self.Q.addConnection(connHidden1ToHidden2)
self.Q.addConnection(connHidden2ToOut)
self.Q.sortModules()
def BackupNetwork(genome=None):
#initial a network [12,12,4] and initial weights are baseline policy versions
from pybrain.structure import FeedForwardNetwork,LinearLayer,TanhLayer,FullConnection
network = FeedForwardNetwork()
inLayer= LinearLayer(12)
hiddenLayer = LinearLayer(12)
outLayer = TanhLayer(4)
network.addInputModule(inLayer)
network.addModule(hiddenLayer)
network.addOutputModule(outLayer)
weights = []
if(genome == None):
import pickle
weights = pickle.load(open("seed"))
else:
weights = genome
in_to_hidden = FullConnection(inLayer,hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer,outLayer)
for i in range(0,144):
in_to_hidden.params[i] = weights[i]
for j in range(0,48):
hidden_to_out.params[j] = weights[j+144]
network.addConnection(in_to_hidden)
network.addConnection(hidden_to_out)
network.sortModules()
return network
def trainFunc(params):
iter, trainds, validds, input_size, hidden, func, eta, lmda, epochs = params
print('Iter:', iter, 'Epochs:', epochs, 'Hidden_size:', hidden, 'Eta:',
eta, 'Lamda:', lmda, 'Activation:', func)
# Build network
n = RecurrentNetwork()
n.addInputModule(LinearLayer(input_size, name='in'))
n.addModule(func(hidden, name='hidden'))
n.addModule(LinearLayer(hidden, name='context'))
n.addOutputModule(LinearLayer(1, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='in_to_hidden'))
n.addConnection(FullConnection(n['hidden'], n['out'],
name='hidden_to_out'))
n.addRecurrentConnection(FullConnection(n['hidden'], n['context']))
rnet = n
rnet.sortModules()
trainer = BackpropTrainer(n,
trainds,
learningrate=eta,
weightdecay=lmda,
momentum=0.1,
shuffle=False)
trainer.trainEpochs(epochs)
pred = np.nan_to_num(n.activateOnDataset(validds))
validerr = eval.calc_RMSE(validds['target'], pred)
varscore = explained_variance_score(validds['target'], pred)
return validerr, varscore, n
def xor_network(self, net):
net.addInputModule(LinearLayer(2, name='in'))
net.addModule(BiasUnit(name='bias'))
net.addModule(LinearLayer(3, name='hidden'))
net.addOutputModule(LinearLayer(1, name='out'))
net.addConnection(FullConnection(net['in'], net['hidden']))
net.addConnection(FullConnection(net['bias'], net['hidden']))
net.addConnection(FullConnection(net['hidden'], net['out']))
def init_network(self, net):
net.addInputModule(LinearLayer(2, 'in'))
net.addModule(SigmoidLayer(3, 'hidden'))
net.addOutputModule(LinearLayer(2, 'out'))
net.addModule(BiasUnit(name='bias'))
net.addConnection(FullConnection(net['in'], net['hidden']))
net.addConnection(FullConnection(net['hidden'], net['out']))
net.sortModules()
def adicionaConexoes(self):
self.ligacao_entrada_oculta = FullConnection(self.camada_entrada,
self.camada_oculta)
self.ligacao_oculta_saida = FullConnection(self.camada_oculta,
self.camada_saida)
self.network.addConnection(self.ligacao_oculta_saida)
self.network.addConnection(self.ligacao_entrada_oculta)
self.iniciaRede()
def fit_predict(xTrain, yTrain, xTest, epochs, neurons):
# Check edge cases
if (not len(xTrain) == len(yTrain) or len(xTrain) == 0 or len(xTest) == 0
or epochs <= 0):
return
# Randomize the training data (probably not necessary but pybrain might
# not shuffle the data itself, so perform as safety check)
indices = np.arange(len(xTrain))
np.random.shuffle(indices)
trainSwapX = [xTrain[x] for x in indices]
trainSwapY = [yTrain[x] for x in indices]
supTrain = SupervisedDataSet(len(xTrain[0]), 1)
for x in range(len(trainSwapX)):
supTrain.addSample(trainSwapX[x], trainSwapY[x])
# Construct the feed-forward neural network
n = FeedForwardNetwork()
inLayer = LinearLayer(len(xTrain[0]))
hiddenLayer1 = SigmoidLayer(neurons)
outLayer = LinearLayer(1)
n.addInputModule(inLayer)
n.addModule(hiddenLayer1)
n.addOutputModule(outLayer)
in_to_hidden = FullConnection(inLayer, hiddenLayer1)
hidden_to_out = FullConnection(hiddenLayer1, outLayer)
n.addConnection(in_to_hidden)
n.addConnection(hidden_to_out)
n.sortModules()
# Train the neural network on the training partition, validating
# the training progress on the validation partition
trainer = BackpropTrainer(n,
dataset=supTrain,
momentum=0.1,
learningrate=0.01,
verbose=False,
weightdecay=0.01)
trainer.trainUntilConvergence(dataset=supTrain,
maxEpochs=epochs,
validationProportion=0.30)
outputs = []
for x in xTest:
outputs.append(n.activate(x))
return outputs
def testMdlstm(self):
net = FeedForwardNetwork()
net.addInputModule(LinearLayer(1, name='in'))
net.addModule(MDLSTMLayer(1, 1, name='hidden'))
net.addOutputModule(LinearLayer(1, name='out'))
net.addConnection(FullConnection(net['in'], net['hidden']))
net.addConnection(FullConnection(net['hidden'], net['out']))
net.sortModules()
self.equivalence_feed_forward(net, net.convertToFastNetwork())
def buildNN(indim=4, hiddim=6, outdim=3):
net = FeedForwardNetwork()
net.addInputModule(TanhLayer(indim, name='i'))
net.addModule(TanhLayer(hiddim, name='h'))
net.addOutputModule(ThresholdLayer(outdim, name='o', threshold=0.5))
net.addConnection(FullConnection(net['i'], net['h']))
net.addConnection(FullConnection(net['h'], net['o']))
net.sortModules()
return net
def __init__(self, x, y, direction):
self.age = 0
# position
self.x = x
self.y = y
# number of fruits peeled
self.num_peeled = 0
self.num_eaten = 0
self.num_moved = 0
# orientation (0 - 359 degrees)
self.direction = direction
# touching anything
self.touching = None
self.sees = None
# hunger sensor
self.hunger = 2000
self.avg_hunger = 0
###
# Neural Network
#
# Inputs:
# 1. sees_peeled_orange
# 2. sees_unpeeled_orange
# 3. sees_peeled_banana
# 4. sees_unpeeled_banana
# 5. sees_animat
# 6. sees_wall
# 7. hunger
# 8. touching_peeled_orange
# 9. touching_unpeeled_orange
# 10. touching_peeled_banana
# 11. touching_unpeeled_banana
# 12. touching_animat
# 13. touching_wall
###
self.net = FeedForwardNetwork()
self.net.addInputModule(LinearLayer(13, name='in'))
self.net.addModule(SigmoidLayer(14, name='hidden'))
self.net.addOutputModule(LinearLayer(5, name='out'))
self.net.addConnection(
FullConnection(self.net['in'], self.net['hidden']))
self.net.addConnection(
FullConnection(self.net['hidden'], self.net['out']))
self.net.sortModules()
# thresholds for deciding an action
self.move_threshold = 0
self.peel_threshold = 0
self.eat_threshold = 0
def getFitness(self, smMatrix): #Store the sm state into memory
fit = 0
#Fitness function (3) *************************************************************
#Record the sm data for this loop and consider its properties
#print(smMatrix)
#print(len(smMatrix))
#net = buildNetwork(3,10,1, bias = True)
net = FeedForwardNetwork()
inp = LinearLayer(3)
h1 = SigmoidLayer(10)
outp = LinearLayer(1)
# add modules
net.addOutputModule(outp)
net.addInputModule(inp)
net.addModule(h1)
# create connections
iToH = FullConnection(inp, h1)
hToO = FullConnection(h1, outp)
net.addConnection(iToH)
net.addConnection(hToO)
# finish up
net.sortModules()
ds = SupervisedDataSet(3, 1)
trainSet = []
for index_x, x in enumerate(smMatrix):
if index_x > 0 and index_x < len(smMatrix) - 1:
#trainSet.append( [smMatrix[index_x][0], smMatrix[index_x][1], smMatrix[index_x][2], smMatrix[index_x+1][3] ] )
ds.addSample(([
smMatrix[index_x][0], smMatrix[index_x][1],
smMatrix[index_x][2]
]), (smMatrix[index_x + 1][3]))
#print(trainSet)
#print(ds)
trainer = BackpropTrainer(net, ds, weightdecay=0.01)
err = trainer.trainUntilConvergence(maxEpochs=100)
#Visualize the network performance and structure.
#nn = NNregression(ds, epoinc = 10)
#nn.setupNN()
#nn.runTraining()
#self.pesos_conexiones(net)
print("Input to hidden", iToH.params)
#print("H to output", hToO.params)
#print(iToH.params)
n1 = iToH.params
n1a = zip(*[iter(n1)] * 3)
n2 = hToO.params
fit = sum(n1a[:]) + sum(n2[:])
print fit
return fit
def runNeuralSimulation(dataTrain, dataTest, train_tfidf, test_tfidf):
outFile = open('neuralLog.txt','a')
outFile.write('-------------------------------------\n')
outFile.write('train==> %d, %d \n'%(train_tfidf.shape[0],train_tfidf.shape[1]))
outFile.write('test==> %d, %d \n'%(test_tfidf.shape[0],test_tfidf.shape[1]))
trainDS = getDataSetFromTfidf(train_tfidf, dataTrain.target)
testDS = getDataSetFromTfidf(test_tfidf, dataTest.target)
print "Number of training patterns: ", len(trainDS)
print "Input and output dimensions: ", trainDS.indim, trainDS.outdim
print "First sample (input, target, class):"
print len(trainDS['input'][0]), trainDS['target'][0], trainDS['class'][0]
# with SimpleTimer('time to train', outFile):
# net = buildNetwork(trainDS.indim, trainDS.indim/2, trainDS.indim/4, trainDS.indim/8, trainDS.indim/16, 2, hiddenclass=TanhLayer, outclass=SoftmaxLayer)
# trainer = BackpropTrainer( net, dataset=trainDS, momentum=0.1, verbose=True, weightdecay=0.01, batchlearning=True)
net = RecurrentNetwork()
net.addInputModule(LinearLayer(trainDS.indim, name='in'))
net.addModule(SigmoidLayer(trainDS.indim/2, name='hidden'))
net.addModule(SigmoidLayer(trainDS.indim/4, name='hidden2'))
net.addOutputModule(SoftmaxLayer(2, name='out'))
net.addConnection(FullConnection(net['in'], net['hidden'], name='c1'))
net.addConnection(FullConnection(net['hidden'], net['out'], name='c2'))
net.addRecurrentConnection(FullConnection(net['hidden'], net['hidden'], name='c3'))
net.addRecurrentConnection(FullConnection(net['hidden2'], net['hidden'], name='c4'))
net.sortModules()
trainer = BackpropTrainer( net, dataset=trainDS, momentum=0.01, verbose=True, weightdecay=0.01)
outFile.write('%s \n' % (net.__str__()))
epochs = 2000
with SimpleTimer('time to train %d epochs' % epochs, outFile):
for i in range(epochs):
trainer.trainEpochs(1)
trnresult = percentError( trainer.testOnClassData(),
trainDS['class'] )
tstresult = percentError( trainer.testOnClassData(
dataset=testDS ), testDS['class'] )
print "epoch: %4d" % trainer.totalepochs, \
" train error: %5.2f%%" % trnresult, \
" test error: %5.2f%%" % tstresult
outFile.write('%5.2f , %5.2f \n' % (100.0-trnresult, 100.0-tstresult))
predicted = trainer.testOnClassData(dataset=testDS)
results = predicted == testDS['class'].flatten()
wrong = []
for i in range(len(results)):
if not results[i]:
wrong.append(i)
print 'classifier got these wrong:'
for i in wrong[:10]:
print dataTest.data[i], dataTest.target[i]
outFile.write('%s %d \n' % (dataTest.data[i], dataTest.target[i]))
def fillConnections(self, net, addedStack, stackToGo, layers):
connections = []
recurrentConnections = []
if len(stackToGo) == 0:
return connections, recurrentConnections
ways = []
futureStack = []
for neuron in stackToGo:
way = []
for w in ways:
if w.count(neuron) != 0:
way = w
else:
continue
if not way:
way.append(neuron)
ways.append(way)
for connection in range(len(net[neuron])):
if net[neuron][connection] == 1:
if addedStack.count(connection) != 0:
recurrentConnections.append(FullConnection(layers[neuron], layers[connection]))
else:
if stackToGo.count(connection) != 0:
if way.count(connection) != 0:
recurrentConnections.append(FullConnection(layers[neuron], layers[connection]))
else:
flag = True
for w in ways:
if w.count(connection) != 0:
connections.append(FullConnection(layers[neuron], layers[connection]))
for n in w:
way.append(n)
ways.pop(ways.index(w))
flag = False
break
else:
continue
if flag:
way.append(connection)
connections.append(FullConnection(layers[neuron], layers[connection]))
else:
connections.append(FullConnection(layers[neuron], layers[connection]))
futureStack.append(connection)
else:
continue
for v in stackToGo:
addedStack.append(v)
c, rc = self.fillConnections(net, addedStack, futureStack, layers)
for con in c:
connections.append(con)
for rcon in rc:
recurrentConnections.append(rcon)
return connections, recurrentConnections
def rec_three_layer_network(self, net):
inlayer = LinearLayer(1, name='in')
hiddenlayer = LinearLayer(1, name='hidden')
outlayer = LinearLayer(1, name='out')
con1 = FullConnection(inlayer, hiddenlayer)
con2 = FullConnection(hiddenlayer, outlayer)
con3 = FullConnection(hiddenlayer, hiddenlayer)
net.addInputModule(inlayer)
net.addModule(hiddenlayer)
net.addOutputModule(outlayer)
net.addConnection(con1)
net.addConnection(con2)
net.addRecurrentConnection(con3)
def lstm_network(self, net):
i = LinearLayer(1, name='in')
h = LSTMLayer(2, name='hidden')
o = LinearLayer(1, name='out')
b = BiasUnit(name='bias')
net.addModule(b)
net.addOutputModule(o)
net.addInputModule(i)
net.addModule(h)
net.addConnection(FullConnection(i, h))
net.addConnection(FullConnection(b, h))
net.addRecurrentConnection(FullConnection(h, h))
net.addConnection(FullConnection(h, o))
def createNNLong(trndata):
nn = FeedForwardNetwork()
inLayer = LinearLayer(trndata.indim, name='in')
hiddenLayer = TanhLayer(6, name='hidden0')
outLayer = TanhLayer(trndata.outdim, name='out')
nn.addInputModule(inLayer)
nn.addModule(hiddenLayer)
nn.addOutputModule(outLayer)
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer, outLayer)
nn.addConnection(in_to_hidden)
nn.addConnection(hidden_to_out)
nn.sortModules()
return nn
def constructNet(self, input, hidden, output):
inputLayer = LinearLayer(input)
hiddenLayer = TanhLayer(hidden)
outputLayer = LinearLayer(output)
self.net.addInputModule(inputLayer)
self.net.addModule(hiddenLayer)
self.net.addOutputModule(outputLayer)
conn1 = FullConnection(inputLayer, hiddenLayer)
conn2 = FullConnection(hiddenLayer, outputLayer)
self.net.addConnection(conn1)
self.net.addConnection(conn2)
def __init__(self, hidden_layers, ally_champ_obj_list,
enemy_champ_obj_list):
self.ally_champ_obj_list = ally_champ_obj_list
self.enemy_champ_obj_list = enemy_champ_obj_list
self.set_nodes()
self.network = FeedForwardNetwork()
connect_queue = Queue.Queue()
for layer in xrange(0, hidden_layers):
connect_queue.put(
TanhLayer(self.input_node_count,
name='hidden_layer_{}'.format(layer)))
connect_queue.put(SigmoidLayer(1, name='output_layer'))
prev_layer = LinearLayer(self.input_node_count, name='input_layer')
self.network.addInputModule(prev_layer)
while not connect_queue.empty():
current_layer = connect_queue.get()
if current_layer.name == 'output_layer':
self.network.addOutputModule(current_layer)
else:
self.network.addModule(current_layer)
bias = BiasUnit()
bias_connection = FullConnection(
bias,
current_layer,
name="bias_to_{}_connection".format(current_layer.name))
self.network.addModule(bias)
self.network.addConnection(bias_connection)
connection = FullConnection(prev_layer,
current_layer,
name="{}_to_{}_connection".format(
prev_layer.name,
current_layer.name))
self.network.addConnection(connection)
prev_layer = current_layer
self.network.sortModules()
def classify(imSize, dataset, hidden_neurons, initial_error):
tstdata, trndata = dataset.splitWithProportion(0.25)
# nos da una proporcion de data de entrenamiento de .75 y prueba .25
# imSize es el tamano de las capas de entrada
# define layer structures
inLayer = LinearLayer(imSize)
hiddenLayer = SigmoidLayer(imSize / 3)
outLayer = SoftmaxLayer(1)
# add layers to network
net = FeedForwardNetwork()
net.addInputModule(inLayer)
net.addModule(hiddenLayer)
net.addOutputModule(outLayer)
# define conncections for network
theta1 = FullConnection(inLayer, hiddenLayer)
theta2 = FullConnection(hiddenLayer, outLayer)
# add connections to network
net.addConnection(theta1)
net.addConnection(theta2)
# sort module
net.sortModules()
dataset._convertToOneOfMany()
fnn = buildNetwork(dataset.indim,
imSize / 3,
dataset.outdim,
outclass=SoftmaxLayer)
#Creamos un entrenador de retropropagacion usando el dataset y la red
trainer = BackpropTrainer(fnn, dataset)
error = initial_error
iteration = 0
#iteramos mientras el error sea menor 0.001
while error > 0.01:
error = trainer.train()
iteration += 1
#print "Iteration: {0} Error {1}".format(iteration, error)
print "Terminado luego de: ", iteration, " iteraciones"
print "Con un error de: ", error
return fnn
def top_layer(self, autoencoder, hidden_layers, next_layer, bias_layers):
# connect 2nd to last and last
last_layer = hidden_layers[-1].outmod
autoencoder.addOutputModule(last_layer)
connection = FullConnection(next_layer, last_layer)
connection.params[:] = hidden_layers[-1].params
autoencoder.addConnection(connection)
if self.bias:
bias = bias_layers[-1]
bias_unit = bias.inmod
autoencoder.addModule(bias_unit)
connection = FullConnection(bias_unit, last_layer)
connection.params[:] = bias.params
autoencoder.addConnection(connection)
autoencoder.sortModules()
return autoencoder
def train(self):
# We will build up a network piecewise in order to create a new dataset
# for each layer.
dataset = self.dataset
piecenet = FeedForwardNetwork()
piecenet.addInputModule(copy.deepcopy(self.net.inmodules[0]))
# Add a bias
bias = BiasUnit()
piecenet.addModule(bias)
# Add the first visible layer
firstRbm = self.iterRbms().next()
visible = copy.deepcopy(firstRbm.visible)
piecenet.addModule(visible)
# For saving the rbms and their inverses
self.invRbms = []
self.rbms = []
for rbm in self.iterRbms():
self.net.sortModules()
# Train the first layer with an rbm trainer for `epoch` epochs.
trainer = self.trainerKlass(rbm, dataset, self.cfg)
for _ in xrange(self.epochs):
trainer.train()
self.invRbms.append(trainer.invRbm)
self.rbms.append(rbm)
# Add the connections and the hidden layer of the rbm to the net.
hidden = copy.deepcopy(rbm.hidden)
biascon = FullConnection(bias, hidden)
biascon.params[:] = rbm.biasWeights
con = FullConnection(visible, hidden)
con.params[:] = rbm.weights
piecenet.addConnection(biascon)
piecenet.addConnection(con)
piecenet.addModule(hidden)
# Overwrite old outputs
piecenet.outmodules = [hidden]
piecenet.outdim = rbm.hiddenDim
piecenet.sortModules()
dataset = UnsupervisedDataSet(rbm.hiddenDim)
for sample, in self.dataset:
new_sample = piecenet.activate(sample)
dataset.addSample(new_sample)
visible = hidden
def fromDims(cls, visibledim, hiddendim, params=None, biasParams=None):
"""Return a restricted Boltzmann machine of the given dimensions with the
given distributions."""
net = FeedForwardNetwork()
bias = BiasUnit('bias')
visible = LinearLayer(visibledim, 'visible')
hidden = SigmoidLayer(hiddendim, 'hidden')
con1 = FullConnection(visible, hidden)
con2 = FullConnection(bias, hidden)
if params is not None:
con1.params[:] = params
if biasParams is not None:
con2.params[:] = biasParams
net.addInputModule(visible)
net.addModule(bias)
net.addOutputModule(hidden)
net.addConnection(con1)
net.addConnection(con2)
net.sortModules()
return cls(net)
def buildNetwork():
# make network objects
network = FeedForwardNetwork()
inputLayer = LinearLayer(N*N)
hiddenLayer = SigmoidLayer(hiddenNodes)
outputLayer = SoftmaxLayer(len(CLASSES))
# connect the network
network.addInputModule(inputLayer)
network.addModule(hiddenLayer)
network.addOutputModule(outputLayer)
input_to_hidden = FullConnection(inputLayer, hiddenLayer)
hidden_to_output = FullConnection(hiddenLayer, outputLayer)
# connection weights determined by training
input_to_hidden._setParameters([4.35074052, 0.54309903, -1.70788914, -0.37641228, 5.36652276, -5.95097706, -3.31479105, 1.48726254, 4.01124973, -2.23954635, -2.06566738, -1.05526604, 0.29454287, 0.34454901, 0.85887803, -1.54283834, -0.20189867, 2.39341244, -6.41137004, -5.21972223, 3.9786469, 2.62502833, -0.71182606, 1.49128153, 4.07595571, 3.84903291, 1.23479208, -1.84971481, -2.48423784, -2.89135785, 1.66872929, 2.9749277, 0.13388845, -6.34083074, 2.45295962, 4.37867807, -2.04605488, 5.83143133, 8.31634908, -1.15218811, -1.67941218, -2.21080881, 0.73068735, -1.38228386, -3.62022672, -0.93999936, 0.93052909, -3.83909343, -4.79640119, 3.39088663, 1.88639523, -2.10136095, -5.79122022, -0.39145108, -3.16506474, -0.99953878, -4.03241107, 2.27154235, 1.29838529, 1.65980538, 3.56765327, 0.2334956, -1.50648655, 3.43253185, 1.96654523, -0.52561201, 2.16708779, 3.43269409, -0.89938682, -4.57967624, -0.54859459, -0.02998314, 0.57531064, -3.22645606, -2.01130185, -0.92961716, 2.72892938, 1.70440582, -0.74359544, -1.75646309])
hidden_to_output._setParameters([5.23726518, 4.23002634, -10.17060945, 0.24449631, -0.40430924, -5.10170453, -1.97718355, -3.29225334, 9.76797256, -0.31218471, -2.26134518, 8.19478967, 4.64314258, -5.10091146, -6.16245232, 5.12979712, -6.02720454, 0.98557444, -7.00566727, 5.45351844, -6.94985525, -5.76529963, 10.43833636, -0.35966021, -0.53525741])
network.addConnection(input_to_hidden)
network.addConnection(hidden_to_output)
network.sortModules() # initialize
return network
def _train(self):
hidden_layers = []
bias_layers = []
compressed_data = copy.copy(self.unsupervised) # it isn't compressed at this point, but will be later on
compressed_supervised = self.supervised
mid_layers = self.layers[1:-1] # remove the first and last
for i,current in enumerate(mid_layers):
prior = self.layers[i] # This accesses the layer before the "current" one, since the indexing in mid_layers and self.layers is offset by 1
# build the NN with a bottleneck
bottleneck = FeedForwardNetwork()
in_layer = LinearLayer(prior)
hidden_layer = self.hidden_layer(current)
out_layer = self.hidden_layer(prior)
bottleneck.addInputModule(in_layer)
bottleneck.addModule(hidden_layer)
bottleneck.addOutputModule(out_layer)
in_to_hidden = FullConnection(in_layer, hidden_layer)
hidden_to_out = FullConnection(hidden_layer, out_layer)
bottleneck.addConnection(in_to_hidden)
bottleneck.addConnection(hidden_to_out)
if self.bias:
bias1 = BiasUnit()
bias2 = BiasUnit()
bottleneck.addModule(bias1)
bottleneck.addModule(bias2)
bias_in = FullConnection(bias1, hidden_layer)
bias_hidden = FullConnection(bias2, out_layer)
bottleneck.addConnection(bias_in)
bottleneck.addConnection(bias_hidden)
bottleneck.sortModules()
# train the bottleneck
print "\n...training for layer ", prior, " to ", current
ds = SupervisedDataSet(prior,prior)
if self.dropout_on:
noisy_data, originals = self.dropout(compressed_data, noise=0.2, bag=1)
for i,n in enumerate(noisy_data):
original = originals[i]
ds.addSample(n, original)
else:
for d in (compressed_data): ds.addSample(d, d)
trainer = BackpropTrainer(bottleneck, dataset=ds, learningrate=0.001, momentum=0.05, verbose=self.verbose, weightdecay=0.05)
trainer.trainEpochs(self.compression_epochs)
if self.verbose: print "...data:\n...", compressed_data[0][:8], "\nreconstructed to:\n...", bottleneck.activate(compressed_data[0])[:8]
hidden_layers.append(in_to_hidden)
if self.bias: bias_layers.append(bias_in)
# use the params from the bottleneck to compress the training data
compressor = FeedForwardNetwork()
compressor.addInputModule(in_layer)
compressor.addOutputModule(hidden_layer) # use the hidden layer from above
compressor.addConnection(in_to_hidden)
compressor.sortModules()
compressed_data = [compressor.activate(d) for d in compressed_data]
compressed_supervised = [compressor.activate(d) for d in compressed_supervised]
self.nn.append(compressor)
# Train the softmax layer
print "\n...training for softmax layer "
softmax = FeedForwardNetwork()
in_layer = LinearLayer(self.layers[-2])
out_layer = self.final_layer(self.layers[-1])
softmax.addInputModule(in_layer)
softmax.addOutputModule(out_layer)
in_to_out = FullConnection(in_layer, out_layer)
softmax.addConnection(in_to_out)
if self.bias:
bias = BiasUnit()
softmax.addModule(bias)
bias_in = FullConnection(bias, out_layer)
softmax.addConnection(bias_in)
softmax.sortModules()
# see if it's for classification or regression
if self.final_layer == SoftmaxLayer:
print "...training for a softmax network"
ds = ClassificationDataSet(self.layers[-2], 1)
else:
print "...training for a regression network"
ds = SupervisedDataSet(self.layers[-2], self.layers[-1])
bag = 1
noisy_data, _ = self.dropout(compressed_supervised, noise=0.5, bag=bag)
bagged_targets = []
for t in self.targets:
for b in range(bag):
bagged_targets.append(t)
for i,d in enumerate(noisy_data):
target = bagged_targets[i]
ds.addSample(d, target)
# see if it's for classification or regression
if self.final_layer == SoftmaxLayer:
ds._convertToOneOfMany()
# TODO make these configurable
trainer = BackpropTrainer(softmax, dataset=ds, learningrate=0.001, momentum=0.05, verbose=self.verbose, weightdecay=0.05)
trainer.trainEpochs(self.compression_epochs)
self.nn.append(softmax)
hidden_layers.append(in_to_out)
if self.bias: bias_layers.append(bias_in)
# Recreate the whole thing
# connect the first two
autoencoder = FeedForwardNetwork()
first_layer = hidden_layers[0].inmod
next_layer = hidden_layers[0].outmod
autoencoder.addInputModule(first_layer)
connection = FullConnection(first_layer, next_layer)
connection.params[:] = hidden_layers[0].params
autoencoder.addConnection(connection)
# decide whether this should be the output layer or not
if self.autoencoding_only and (len(self.layers) <= 3): # TODO change this to 2 when you aren't using the softmax above
autoencoder.addOutputModule(next_layer)
else:
autoencoder.addModule(next_layer)
if self.bias:
bias = bias_layers[0]
bias_unit = bias.inmod
autoencoder.addModule(bias_unit)
connection = FullConnection(bias_unit, next_layer)
connection.params[:] = bias.params
autoencoder.addConnection(connection)
# connect the middle layers
for i,h in enumerate(hidden_layers[1:-1]):
new_next_layer = h.outmod
# decide whether this should be the output layer or not
if self.autoencoding_only and i == (len(hidden_layers) - 3):
autoencoder.addOutputModule(new_next_layer)
else:
autoencoder.addModule(new_next_layer)
connection = FullConnection(next_layer, new_next_layer)
connection.params[:] = h.params
autoencoder.addConnection(connection)
next_layer = new_next_layer
if self.bias:
bias = bias_layers[i+1]
bias_unit = bias.inmod
autoencoder.addModule(bias_unit)
connection = FullConnection(bias_unit, next_layer)
connection.params[:] = bias.params
autoencoder.addConnection(connection)
return autoencoder, hidden_layers, next_layer, bias_layers
def simple_network_builder(layers,partial_path):
n = FeedForwardNetwork()
## create the network
inlayer = LinearLayer(layers[0], name = "In")
hidden_one = TanhLayer(layers[1], name = "Hidden 1")
hidden_two = TanhLayer(layers[2], name ="Hidden 2")
b1 = BiasUnit(name="Bias")
output = LinearLayer(1,name = "Out")
n.addInputModule(inlayer)
n.addModule(hidden_one)
n.addModule(hidden_two)
n.addModule(b1)
n.addOutputModule(output)
in_to_one = FullConnection(inlayer,hidden_one)
one_to_two = FullConnection(hidden_one,hidden_two)
two_to_out = FullConnection(hidden_two,output)
b1_to_one = FullConnection(b1,hidden_one)
b2_to_two = FullConnection(b1,hidden_two)
b3_to_output = FullConnection(b1,output)
### load weights and biases
in_to_one._setParameters(np.array((csv_loader(partial_path + '_w1.csv'))))
one_to_two._setParameters(np.array(csv_loader(partial_path + '_w2.csv')))
two_to_out._setParameters(np.array(csv_loader(partial_path + '_w3.csv')))
b1_to_one._setParameters(np.array(csv_loader(partial_path + '_b1.csv')))
b2_to_two._setParameters(np.array(csv_loader(partial_path + '_b2.csv')))
b3_to_output._setParameters(np.array(csv_loader(partial_path + '_b3.csv')))
### connect the network topology
n.addConnection(in_to_one)
n.addConnection(one_to_two)
n.addConnection(two_to_out)
# n.sortModules()
n.addConnection(b1_to_one)
n.addConnection(b2_to_two)
n.addConnection(b3_to_output)
### finalize network object
n.sortModules()
return n
def StateToActionNetwork():
#initial a network [12,12,4] and initial weights are baseline policy versions
from pybrain.structure import FeedForwardNetwork,LinearLayer,TanhLayer,FullConnection
network = FeedForwardNetwork()
inLayer= LinearLayer(12)
hiddenLayer = TanhLayer(12)
outLayer = TanhLayer(4)
network.addInputModule(inLayer)
network.addModule(hiddenLayer)
network.addOutputModule(outLayer)
#inital weights
initial_weights = """-0.82254189 -1.13179445 -0.05073786 0.26591425 -0.1284119 0.08943027
0.42276271 -0.7071644 -1.45276617 -0.44496227 0.59200697 -0.76490859
-0.82167338 -0.39902595 -0.34932747 -1.5301132 0.4874284 -1.75511689
0.48486169 0.81363237 -0.15560306 -1.28014402 0.432026 -0.245171
0.78031838 -0.15817382 0.11117014 1.04968207 -0.2928946 -1.83646693
1.44371163 -0.16511239 -0.28240856 0.66388542 -1.82 -0.31922762
-1.62204838 -0.12312791 0.22280484 0.7411014 -0.05863777 -0.5328774
-2.78973853 0.46491466 1.42202806 -2.93244732 -0.24784862 -1.09437463
-0.07650338 -0.22306763 -0.0183736 1.78186929 -0.67513165 -0.55366751
0.5925835 -1.82412031 -0.05014905 -0.53091446 -1.00910792 1.35670824
1.58071742 -1.65940386 1.17890985 -1.64222732 0.56357161 -0.304062
0.83628703 0.80113178 -0.09179465 0.74009935 -0.34146348 0.45395903
-0.80083394 -0.3178608 -0.46622104 -1.59866551 -0.3893681 -0.67853351
-0.61304091 -0.12200701 1.65532154 -0.33992727 -1.56087088 -2.02031568
1.02997029 1.21253299 -0.06733012 0.28724485 -0.27014336 -0.83057191
0.39323538 1.30558669 0.26726448 -0.65961121 1.54584633 0.09210854
-0.99081429 0.59634696 0.08429763 -0.1085911 0.01785386 1.78681155
0.87636657 -1.2413246 0.08531575 -0.92648945 -0.20240477 0.16277405
-0.23555818 0.49663751 2.75629226 0.03482599 -0.64754342 -0.61886618
0.77400906 -0.28588506 -0.18226857 -1.15349435 0.09339758 -0.84021149
-0.3769615 -2.71952741 -1.92806955 1.33770349 0.46549708 1.0234573
0.77816064 -0.36149316 0.65660944 -0.79934234 -0.85783489 -0.10840895
1.35537789 -1.50803792 0.10239295 0.20335467 0.07891178 -0.56889871
0.59446914 -1.07917626 -1.44565869 0.46396979 -1.0022648 -0.36037274
0.04604105 0.10828613 -0.09156346 0.05961271 0.07350161 0.03483664
0.01918546 0.029282 -0.03780433 0.01140018 -0.04217829 0.12422228
0.10494436 0.03090324 0.02751887 0.1757922 -0.1175768 0.04984245
0.03805592 0.07699565 0.0927753 0.03017363 -0.02785207 -0.08504634
0.23548627 -0.024849 0.08893206 -0.02284833 0.04222917 -0.01530065
-0.0336135 0.08849411 -0.02291273 -0.05779803 -0.01868145 0.00836078
0.08720535 -0.11581814 -0.03772317 0.05162675 0.10993543 0.08677515
-0.03086664 -0.02367544 0.10032227 0.15426584 -0.03793561 -0.07125042"""
weights = []
for i in initial_weights.split(" "):
num = i.strip().replace('\t','')
num = num.replace('\n',' ')
num = num.strip()
try:
weights.append(float(num))
except:
pass
in_to_hidden = FullConnection(inLayer,hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer,outLayer)
for i in range(0,144):
in_to_hidden.params[i] = weights[i]
for j in range(0,48):
hidden_to_out.params[j] = weights[j+144]
network.addConnection(in_to_hidden)
network.addConnection(hidden_to_out)
network.sortModules()
return network
def _train(self):
global bias_in
hidden_layers = []
bias_layers = []
compressed_data = copy.copy(self.unsupervised) # it isn't compressed at this point, but will be later on
compressed_supervised = self.supervised
mid_layers = self.layers[1:-1] # remove the first and last
for i, current in enumerate(mid_layers):
prior = self.layers[i]
# This accesses the layer before the "current" one,
# since the indexing in mid_layers and self.layers is offset by 1
# print "Compressed data at stage {0} {1}".format(i, compressed_data)
""" build the NN with a bottleneck """
bottleneck = FeedForwardNetwork()
in_layer = LinearLayer(prior)
hidden_layer = self.hidden_layer(current)
out_layer = self.hidden_layer(prior)
bottleneck.addInputModule(in_layer)
bottleneck.addModule(hidden_layer)
bottleneck.addOutputModule(out_layer)
in_to_hidden = FullConnection(in_layer, hidden_layer)
hidden_to_out = FullConnection(hidden_layer, out_layer)
bottleneck.addConnection(in_to_hidden)
bottleneck.addConnection(hidden_to_out)
if self.bias:
bias1 = BiasUnit()
bias2 = BiasUnit()
bottleneck.addModule(bias1)
bottleneck.addModule(bias2)
bias_in = FullConnection(bias1, hidden_layer)
bias_hidden = FullConnection(bias2, out_layer)
bottleneck.addConnection(bias_in)
bottleneck.addConnection(bias_hidden)
bottleneck.sortModules()
print("3here network is okay bottleneck") # ====================================
""" train the bottleneck """
print "\n...training for layer ", prior, " to ", current
ds = SupervisedDataSet(prior, prior)
print ("5here supervised dataset was built") # ==============================
print("8.====================compressed_data_size=============")
print compressed_data.__sizeof__()
if self.dropout_on:
noisy_data, originals = self.dropout(compressed_data, noise=0.2, bag=1, debug=False)
print("6here dropout is begin processing and it's okay") # ==============================
print "=============noisylen================"
print len(noisy_data) # =====
for i, n in enumerate(noisy_data):
original = originals[i]
ds.addSample(n, original)
print("7.drop out add nosizy sample success") # =============================
else:
for d in (compressed_data):
ds.addSample(d, d)
print("4here begin bp bp bp") # ============================================
trainer = BackpropTrainer(bottleneck, dataset=ds, learningrate=0.001, momentum=0.05,
verbose=self.verbose, weightdecay=0.05)
trainer.trainEpochs(self.compression_epochs)
if self.verbose:
print "...data:\n...", compressed_data[0][:10], \
"\nreconstructed to:\n...", bottleneck.activate(compressed_data[0])[:10]
# just used 10dim of 95 dim mfcc
hidden_layers.append(in_to_hidden)
if self.bias: bias_layers.append(bias_in)
""" use the params from the bottleneck to compress the training data """
compressor = FeedForwardNetwork()
compressor.addInputModule(in_layer)
compressor.addOutputModule(hidden_layer) # use the hidden layer from above
compressor.addConnection(in_to_hidden)
compressor.sortModules()
compressed_data = [compressor.activate(d) for d in compressed_data]
# del compressed_data #del==============================================
compressed_supervised = [compressor.activate(d) for d in compressed_supervised]
# del compressed_supervised #del==============================================
self.nn.append(compressor)
""" Train the softmax layer """
print "\n...training for softmax layer "
softmax = FeedForwardNetwork()
in_layer = LinearLayer(self.layers[-2])
out_layer = self.final_layer(self.layers[-1])
softmax.addInputModule(in_layer)
softmax.addOutputModule(out_layer)
in_to_out = FullConnection(in_layer, out_layer)
softmax.addConnection(in_to_out)
if self.bias:
bias = BiasUnit()
softmax.addModule(bias)
bias_in = FullConnection(bias, out_layer)
softmax.addConnection(bias_in)
softmax.sortModules()
# see if it's for classification or regression
if self.final_layer == SoftmaxLayer:
print "...training for a softmax network"
ds = ClassificationDataSet(self.layers[-2], 1)
else:
print "...training for a regression network"
ds = SupervisedDataSet(self.layers[-2], self.layers[-1])
bag = 1
noisy_data, _ = self.dropout(compressed_supervised, noise=0.5, bag=bag, debug=True)
bagged_targets = []
for t in self.targets:
for b in range(bag):
bagged_targets.append(t)
for i, d in enumerate(noisy_data):
target = bagged_targets[i]
ds.addSample(d, target)
# see if it's for classification or regression
if self.final_layer == SoftmaxLayer:
ds._convertToOneOfMany()
trainer = BackpropTrainer(softmax, dataset=ds, learningrate=0.001,
momentum=0.05, verbose=self.verbose, weightdecay=0.05)
trainer.trainEpochs(self.compression_epochs)
self.nn.append(softmax)
# print "ABOUT TO APPEND"
# print len(in_to_out.params)
hidden_layers.append(in_to_out)
if self.bias:
bias_layers.append(bias_in)
""" Recreate the whole thing """
# print "hidden layers: " + str(hidden_layers)
# print "bias layers: " + str(bias_layers)
# print "len hidden layers: " + str(len(hidden_layers))
# print "len bias layers: " + str(len(bias_layers))
# connect the first two
autoencoder = FeedForwardNetwork()
first_layer = hidden_layers[0].inmod
next_layer = hidden_layers[0].outmod
autoencoder.addInputModule(first_layer)
connection = FullConnection(first_layer, next_layer)
connection.params[:] = hidden_layers[0].params
autoencoder.addConnection(connection)
# decide whether this should be the output layer or not
if self.autoencoding_only and (len(self.layers) <= 3):
# TODO change this to 2 when you aren't using the softmax above
autoencoder.addOutputModule(next_layer)
else:
autoencoder.addModule(next_layer)
if self.bias:
bias = bias_layers[0]
bias_unit = bias.inmod
autoencoder.addModule(bias_unit)
connection = FullConnection(bias_unit, next_layer)
# print bias.params
connection.params[:] = bias.params
autoencoder.addConnection(connection)
# print connection.params
# connect the middle layers
for i, h in enumerate(hidden_layers[1:-1]):
new_next_layer = h.outmod
# decide whether this should be the output layer or not
if self.autoencoding_only and i == (len(hidden_layers) - 3):
autoencoder.addOutputModule(new_next_layer)
else:
autoencoder.addModule(new_next_layer)
connection = FullConnection(next_layer, new_next_layer)
connection.params[:] = h.params
autoencoder.addConnection(connection)
next_layer = new_next_layer
if self.bias:
bias = bias_layers[i + 1]
bias_unit = bias.inmod
autoencoder.addModule(bias_unit)
connection = FullConnection(bias_unit, next_layer)
connection.params[:] = bias.params
autoencoder.addConnection(connection)
return autoencoder, hidden_layers, next_layer, bias_layers