def buildNN(self, net, functions, inp, out):
layers = []
inLayer = self.func[functions[0]](inp)
layers.append(inLayer)
outLayer = self.func[functions[-1]](out)
for neural in range(1, len(net) - 1):
layers.append(self.func[functions[neural]](1))
layers.append(outLayer)
connections, recConnections = self.fillConnections(net, [], [0], layers)
if len(recConnections) == 0:
n = FeedForwardNetwork()
else:
n = RecurrentNetwork()
n.addInputModule(inLayer)
for layer in range(1, len(layers) - 1):
n.addModule(layers[layer])
n.addOutputModule(outLayer)
for con in connections:
n.addConnection(con)
for rcon in recConnections:
n.addRecurrentConnection(rcon)
n.sortModules()
return n
def __init__(self, nin, nout):
singleton.append(self)
self.inn = nin
self.outn = nout
self.n = buildNetwork(nin, 20, nout, bias=False, recurrent=True)
self.n = RecurrentNetwork()
self.n.addInputModule(LinearLayer(nin, name='in'))
self.n.addOutputModule(LinearLayer(nout, name='out'))
self.n.addModule(SigmoidLayer(8, name='hidden2'))
self.n.addModule(TanhLayer(nin+nout/2, name='hidden1'))
self.n.addModule(BiasUnit(name='bias'))
self.n.addModule(LSTMLayer(5, name='memory'))
self.n.addConnection(FullConnection(self.n['in'], self.n['hidden1']))
self.n.addConnection(FullConnection(self.n['bias'], self.n['hidden1']))
self.n.addConnection(FullConnection(self.n['hidden1'], self.n['hidden2']))
self.n.addConnection(FullConnection(self.n['hidden2'], self.n['out']))
self.n.addConnection(FullConnection(self.n['hidden1'], self.n['memory']))
self.n.addConnection(FullConnection(self.n['memory'], self.n['hidden2']))
self.n.addConnection(FullConnection(self.n['in'], self.n['hidden2']))
self.n.addConnection(FullConnection(self.n['hidden2'], self.n['out']))
self.n.addRecurrentConnection(FullConnection(self.n['hidden1'], self.n['hidden1']))
self.n.addRecurrentConnection(FullConnection(self.n['memory'], self.n['hidden1']))
self.n.sortModules()
def __init__(self, dire, x, y, genome):
self.direction = dire
self.x = x
self.y = y
self.genome = genome
self.hunger = 150
self.boredom = 150
self.pain = 0
self.last_hunger = 150
self.last_boredom = 150
self.last_pain = 0
self.last_obj = None
#self.net = FeedForwardNetwork()
self.net = RecurrentNetwork()
self.net.sequential = False
def __init__(self):
self.n = RecurrentNetwork()
inLayer = LinearLayer(8)
hiddenLayer = SigmoidLayer(4)
self.numInputs = 8
outLayer = LinearLayer(4)
self.n.addInputModule(inLayer)
self.n.addModule(hiddenLayer)
self.n.addOutputModule(outLayer)
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer, outLayer)
self.n.addConnection(in_to_hidden)
self.n.addConnection(hidden_to_out)
self.n.sortModules()
self.ds = SupervisedDataSet(8, 4)
self.trainer = BackpropTrainer(self.n, self.ds)
def initialize(self, verbose):
print("Initializing language learner...")
self.verbose = verbose
# Create network and modules
self.net = RecurrentNetwork()
inp = LinearLayer(self.inputs, name="in")
hiddenModules = []
for i in range(0, self.hiddenLayers):
hiddenModules.append(LSTMLayer(self.hiddenNodes, name=("hidden-" + str(i + 1))))
outp = LinearLayer(self.outputs, name="out")
# Add modules to the network with recurrence
self.net.addOutputModule(outp)
self.net.addInputModule(inp)
for module in hiddenModules:
self.net.addModule(module)
# Create connections
self.net.addConnection(FullConnection(self.net["in"], self.net["hidden-1"]))
for i in range(0, len(hiddenModules) - 1):
self.net.addConnection(FullConnection(self.net["hidden-" + str(i + 1)], self.net["hidden-" + str(i + 2)]))
self.net.addRecurrentConnection(FullConnection(self.net["hidden-" + str(i + 1)], self.net["hidden-" + str(i + 1)]))
self.net.addRecurrentConnection(FullConnection(self.net["hidden-" + str(len(hiddenModules))],
self.net["hidden-" + str(len(hiddenModules))]))
self.net.addConnection(FullConnection(self.net["hidden-" + str(len(hiddenModules))], self.net["out"]))
self.net.sortModules()
self.trainingSet = SequentialDataSet(self.inputs, self.outputs)
for x, y in zip(self.dataIn, self.dataOut):
self.trainingSet.newSequence()
self.trainingSet.appendLinked([x], [y])
self.net.randomize()
print("Neural network initialzed with structure:")
print(self.net)
self.trainer = BackpropTrainer(self.net, self.trainingSet, verbose=verbose)
self.__initialized = True
print("Successfully initialized network.")
class BrainController:
indim = 2
outdim = 2
def __init__(self, trained_net=None):
if trained_net == None:
self.net = RecurrentNetwork()
self.init_network(self.net)
else:
self.net = trained_net
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 train(self, data):
ds = SupervisedDataSet(2, 2)
for i in range(0, len(data)):
input, target = data[i]
ds.addSample(input, target)
trainer = BackpropTrainer(self.net,
ds,
learningrate=0.01,
momentum=0.99,
verbose=True)
max_error = 1e-5
error = 1
while abs(error) >= max_error:
error = trainer.train()
#self.validate_net()
f = open('neuro.net', 'w')
pickle.dump(self.net, f)
f.close()
def validate_net(self):
print self.net.activate([0, 0])
print self.net.activate([0, 1])
print self.net.activate([0, 2])
print self.net.activate([1, 0])
print self.net.activate([1, 1])
print self.net.activate([1, 2])
def __init__(self, name, dataset, trained, store):
self.name = name
self.store = store
self.trained = trained
self.dataset = dataset
self.net = RecurrentNetwork()
self.net.addInputModule(LinearLayer(2, name='in'))
self.net.addModule(SigmoidLayer(3, name='hidden'))
self.net.addOutputModule(LinearLayer(2, name='out'))
self.net.addConnection(FullConnection(self.net['in'], self.net['out'], name='c1'))
self.net.addConnection(FullConnection(self.net['hidden'], self.net['out'], name='c2'))
self.net.addRecurrentConnection(FullConnection(self.net['hidden'], self.net['hidden'], name='c3'))
self.net.sortModules()
'''
self.net = buildNetwork(2, 3, 2)
'''
if not self.trained:
self.train()
return
class BrainController:
indim = 2
outdim = 2
def __init__(self, trained_net = None):
if trained_net == None:
self.net = RecurrentNetwork()
self.init_network(self.net)
else:
self.net = trained_net
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 train(self, data):
ds = SupervisedDataSet(2, 2)
for i in range(0, len(data)):
input, target = data[i]
ds.addSample(input, target)
trainer = BackpropTrainer(self.net, ds, learningrate=0.01, momentum=0.99,
verbose=True)
max_error = 1e-5
error = 1
while abs(error) >= max_error:
error = trainer.train()
#self.validate_net()
f = open('neuro.net', 'w')
pickle.dump(self.net, f)
f.close()
def validate_net(self):
print self.net.activate([0, 0])
print self.net.activate([0, 1])
print self.net.activate([0, 2])
print self.net.activate([1, 0])
print self.net.activate([1, 1])
print self.net.activate([1, 2])
def testTraining():
# the AnBnCn dataset (sequential)
d = AnBnCnDataSet()
# build a recurrent network to be trained
hsize = 2
n = RecurrentNetwork()
n.addModule(TanhLayer(hsize, name = 'h'))
n.addModule(BiasUnit(name = 'bias'))
n.addOutputModule(LinearLayer(1, name = 'out'))
n.addConnection(FullConnection(n['bias'], n['h']))
n.addConnection(FullConnection(n['h'], n['out']))
n.addRecurrentConnection(FullConnection(n['h'], n['h']))
n.sortModules()
# initialize the backprop trainer and train
t = BackpropTrainer(n, learningrate = 0.1, momentum = 0.0, verbose = True)
t.trainOnDataset(d, 200)
# the resulting weights are in the network:
print 'Final weights:', n.params
# which are raw price movements)
dataset, tgt = dtools.gen_ds(ltc, 1, ltc_opts, "CRT")
# initialize a pybrain dataset
DS = SupervisedDataSet(len(dataset.values[0]), np.size(tgt.values[0]))
# fill it
for i in xrange(len(dataset)):
DS.appendLinked(dataset.values[i], [tgt.values[i]])
# split 70% for training, 30% for testing
train_set, test_set = DS.splitWithProportion(0.7)
# build our recurrent network with 10 hidden neurodes, one recurrent
# connection, using tanh activation functions
net = RecurrentNetwork()
hidden_neurodes = 10
net.addInputModule(LinearLayer(len(train_set["input"][0]), name="in"))
net.addModule(TanhLayer(hidden_neurodes, name="hidden1"))
net.addOutputModule(LinearLayer(len(train_set["target"][0]), name="out"))
net.addConnection(FullConnection(net["in"], net["hidden1"], name="c1"))
net.addConnection(FullConnection(net["hidden1"], net["out"], name="c2"))
net.addRecurrentConnection(FullConnection(net["out"], net["hidden1"], name="cout"))
net.sortModules()
net.randomize()
# train for 30 epochs (overkill) using the rprop- training algorithm
trainer = RPropMinusTrainer(net, dataset=train_set, verbose=True)
trainer.trainOnDataset(train_set, 30)
# test on training set
from pybrain.structure import RecurrentNetwork, FullConnection from pybrain.structure import LinearLayer # from pybrain.structure import SigmoidLayer from pybrain.structure import TanhLayer from pybrain.datasets import SupervisedDataSet from pybrain.supervised.trainers import BackpropTrainer # Define network structure network = RecurrentNetwork(name="XOR") inputLayer = LinearLayer(2, name="Input") hiddenLayer = TanhLayer(3, name="Hidden") outputLayer = LinearLayer(1, name="Output") network.addInputModule(inputLayer) network.addModule(hiddenLayer) network.addOutputModule(outputLayer) c1 = FullConnection(inputLayer, hiddenLayer, name="Input_to_Hidden") c2 = FullConnection(hiddenLayer, outputLayer, name="Hidden_to_Output") c3 = FullConnection(hiddenLayer, hiddenLayer, name="Recurrent_Connection") network.addConnection(c1) network.addRecurrentConnection(c3) network.addConnection(c2) network.sortModules() # Add a data set
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 trainedRFCNN():
n = RecurrentNetwork()
n.addInputModule(LinearLayer(4, name='in'))
n.addModule(SigmoidLayer(6, name='hidden'))
n.addOutputModule(LinearLayer(2, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(FullConnection(n['out'], n['hidden'], name='nmc'))
n.sortModules()
draw_connections(n)
# d = generateTraininqgData()
d = getDatasetFromFile(root.path()+"/res/dataSet")
t = BackpropTrainer(n, d, learningrate=0.001, momentum=0.75)
t.trainOnDataset(d)
# FIXME: I'm not sure the recurrent ANN is going to converge
# so just training for fixed number of epochs
count = 0
while True:
globErr = t.train()
print globErr
if globErr < 0.01:
break
count = count + 1
if (count == 100):
return trainedRFCNN()
# for i in range(100):
# print t.train()
exportRFCNN(n)
draw_connections(n)
return n
def trainedRFCNN():
n = RecurrentNetwork()
n.addInputModule(LinearLayer(4, name='in'))
n.addModule(SigmoidLayer(6, name='hidden'))
n.addOutputModule(LinearLayer(2, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(FullConnection(n['out'], n['hidden'], name='nmc'))
n.sortModules()
draw_connections(n)
# d = generateTraininqgData()
d = getDatasetFromFile(root.path() + "/res/dataSet")
t = BackpropTrainer(n, d, learningrate=0.001, momentum=0.75)
t.trainOnDataset(d)
# FIXME: I'm not sure the recurrent ANN is going to converge
# so just training for fixed number of epochs
count = 0
while True:
globErr = t.train()
print globErr
if globErr < 0.01:
break
count = count + 1
if (count == 100):
return trainedRFCNN()
# for i in range(100):
# print t.train()
exportRFCNN(n)
draw_connections(n)
return n
def buildMixedNestedNetwork():
""" build a nested network with the inner one being a ffn and the outer one being recurrent. """
N = RecurrentNetwork('outer')
a = LinearLayer(1, name='a')
b = LinearLayer(2, name='b')
c = buildNetwork(2, 3, 1)
c.name = 'inner'
N.addInputModule(a)
N.addModule(c)
N.addOutputModule(b)
N.addConnection(FullConnection(a, b))
N.addConnection(FullConnection(b, c))
N.addRecurrentConnection(FullConnection(c, c))
N.sortModules()
return N
from pybrain.structure import FullConnection from pybrain.supervised.trainers import BackpropTrainer from pybrain.datasets import SupervisedDataSet from pybrain.tools.xml.networkwriter import NetworkWriter from pybrain.utilities import percentError import data_parsing from scipy import array,where training_data = [] ds = SupervisedDataSet(28,39) training_data = data_parsing.conversion_to_one_hot_representation() ds = data_parsing.conversion_to_pybrain_dataset_format(training_data) test, train = ds.splitWithProportion( 0.25 ) n = RecurrentNetwork() input1 = LinearLayer(28) hidden1 = LSTMLayer(512) hidden2 = LSTMLayer(512) hidden3 = LSTMLayer(128) output1 = SigmoidLayer(39) output2 = LinearLayer(39) n.addInputModule(input1) n.addModule(hidden1) n.addModule(hidden2) n.addModule(hidden3) n.addModule(output1) n.addOutputModule(output2)
def buildSimpleLSTMNetwork(peepholes=False):
N = RecurrentNetwork('simpleLstmNet')
i = LinearLayer(100, name='i')
h = LSTMLayer(10, peepholes=peepholes, name='lstm')
o = LinearLayer(1, name='o')
b = BiasUnit('bias')
N.addModule(b)
N.addOutputModule(o)
N.addInputModule(i)
N.addModule(h)
N.addConnection(FullConnection(i, h, name='f1'))
N.addConnection(FullConnection(b, h, name='f2'))
N.addRecurrentConnection(FullConnection(h, h, name='r1'))
N.addConnection(FullConnection(h, o, name='r1'))
N.sortModules()
return N
# which are raw price movements)
dataset, tgt = dtools.gen_ds(ltc, 1, ltc_opts, "CRT")
# initialize a pybrain dataset
DS = SupervisedDataSet(len(dataset.values[0]), np.size(tgt.values[0]))
# fill it
for i in xrange(len(dataset)):
DS.appendLinked(dataset.values[i], [tgt.values[i]])
# split 70% for training, 30% for testing
train_set, test_set = DS.splitWithProportion(.7)
# build our recurrent network with 10 hidden neurodes, one recurrent
# connection, using tanh activation functions
net = RecurrentNetwork()
hidden_neurodes = 10
net.addInputModule(LinearLayer(len(train_set["input"][0]), name="in"))
net.addModule(TanhLayer(hidden_neurodes, name="hidden1"))
net.addOutputModule(LinearLayer(len(train_set["target"][0]), name="out"))
net.addConnection(FullConnection(net["in"], net["hidden1"], name="c1"))
net.addConnection(FullConnection(net["hidden1"], net["out"], name="c2"))
net.addRecurrentConnection(
FullConnection(net["out"], net["hidden1"], name="cout"))
net.sortModules()
net.randomize()
# train for 30 epochs (overkill) using the rprop- training algorithm
trainer = RPropMinusTrainer(net, dataset=train_set, verbose=True)
trainer.trainOnDataset(train_set, 30)
def createJeffersonMDLNetwork(mdl_length=2,
hidden_count=5,
output_count=4,
in_to_out_connect=True,
name=None):
ret_net = RecurrentNetwork(name=name)
# Add some components of the neural network.
hidden_layer = SigmoidLayer(hidden_count, name="hidden")
output_layer = LinearLayer(output_count, name="move")
ret_net.addModule(hidden_layer)
ret_net.addOutputModule(output_layer)
ret_net.addConnection(
FullConnection(hidden_layer,
output_layer,
name="Hidden to Move Layer"))
mdl_prev = ()
for idx in range(0, mdl_length):
# Create the layers
food_layer = LinearLayer(2, name="Food {0}".format(idx))
mdl_layer = LinearLayer(2, name="MDL Layer {0}".format(idx))
# Add to network
ret_net.addModule(food_layer)
if idx == 0:
ret_net.addInputModule(mdl_layer)
else:
ret_net.addModule(mdl_layer)
# Add delay line connection.
ret_net.addRecurrentConnection(
FullConnection(mdl_prev,
mdl_layer,
name="Recurrent DL {0} to DL {1}".format(
idx - 1, idx)))
# Add connections for
# - Delay line to NN.
# - NN to Hidden.
# - NN to Out (if desired).
ret_net.addConnection(
FullConnection(mdl_layer,
food_layer,
name="DL {0} to Food {0}".format(idx)))
ret_net.addConnection(
FullConnection(food_layer,
hidden_layer,
name="Food {0} to Hidden".format(idx)))
if in_to_out_connect:
ret_net.addConnection(
FullConnection(food_layer,
output_layer,
name="Food {0} to Output".format(idx)))
mdl_prev = mdl_layer
ret_net.sortModules()
return ret_net
def createJeffersonStyleNetwork(in_count=2,
hidden_count=5,
output_count=4,
recurrent=True,
in_to_out_connect=True,
name=None):
"""
Creates a Jefferson-esque neural network for trail problem.
Returns:
pybrain.network. The neural network.
"""
if recurrent:
ret_net = RecurrentNetwork(name=name)
else:
ret_net = FeedForwardNetwork(name=name)
in_layer = LinearLayer(in_count, name="food")
hidden_layer = SigmoidLayer(hidden_count, name="hidden")
output_layer = LinearLayer(output_count, name="move")
ret_net.addInputModule(in_layer)
ret_net.addModule(hidden_layer)
ret_net.addOutputModule(output_layer)
in_to_hidden = FullConnection(in_layer, hidden_layer)
hidden_to_out = FullConnection(hidden_layer, output_layer)
ret_net.addConnection(in_to_hidden)
ret_net.addConnection(hidden_to_out)
if in_to_out_connect:
in_to_out = FullConnection(in_layer, output_layer)
ret_net.addConnection(in_to_out)
if recurrent:
hidden_to_hidden = FullConnection(hidden_layer, hidden_layer)
ret_net.addRecurrentConnection(hidden_to_hidden)
ret_net.sortModules()
return ret_net
import itertools
from pybrain.structure import RecurrentNetwork
from pybrain.supervised.trainers import BackpropTrainer
from pybrain.datasets import SupervisedDataSet
john = 1
bill = 2
sue = 3
mary = 4
love = 10
see = 11
ds = SupervisedDataSet(2, 1)
for verb in [love, see]:
for a, b in itertools.combinations([john, bill, sue, mary]):
ds.addSample((verb, a, b), (1, ))
n = RecurrentNetwork()
n.addInputModule(LinearLayer(3, name='in'))
n.addModule(SigmoidLayer(3, name='hidden'))
n.addOutputModule(LinearLayer(1, 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'))
trainer = BackpropTrainer(n, ds)
trainer.train()
def __init__(self, trained_net=None):
if trained_net == None:
self.net = RecurrentNetwork()
self.init_network(self.net)
else:
self.net = trained_net
def main():
numIterations=200
terminal_EMA_SharpeRatio=[0 for i in range(numIterations)]
numTrades=[0 for i in range(numIterations)]
sharpe_first_half=[0 for i in range(numIterations)]
sharpe_sec_half=[0 for i in range(numIterations)]
sharpe_ratio_total=[0 for i in range(numIterations)]
for i in range(numIterations):
env=RWEnvironment(2000)
task = MaximizeReturnTask(env)
numIn=min(env.worldState.shape)
net=RecurrentNetwork()
net.addInputModule(BiasUnit(name='bias'))
net.addOutputModule((SignLayer(1,name='out')))
net.addRecurrentConnection(FullConnection(net['out'], net['out'], name='c3'))
net.addInputModule(LinearLayer(numIn,name='in'))
net.addConnection(FullConnection(net['in'],net['out'],name='c1'))
net.addConnection((FullConnection(net['bias'],net['out'],name='c2')))
net.sortModules()
ts=env.ts
learner = RRL(numIn+2,ts) # ENAC() #Q_LinFA(2,1)
agent = LearningAgent(net,learner)
exp = ContinuousExperiment(task,agent)
#performance tracking
exp.doInteractionsAndLearn(len(ts)-1)
#print(net._params)
terminal_EMA_SharpeRatio[i]=learner.ema_sharpeRatio[-1]
rs=pE.calculateTradingReturn(env.actionHistory,ts)
sharpe_first_half[i]=pE.annualisedSharpe(rs[:(len(ts)/2)])
sharpe_sec_half[i]=pE.annualisedSharpe(rs[len(ts)/2:])
sharpe_ratio_total[i]=pE.annualisedSharpe(rs)
numTrades[i]=learner.numTrades
print(net._params)
print("average number of trades per 1000 observations is {}".format(mean(numTrades)/2))
print("mean Sharpe ratios are {} with standard errors {}, and {} with standard errors {}".format(mean(sharpe_first_half),st.sem(sharpe_first_half),mean(sharpe_sec_half),st.sem(sharpe_sec_half)))
print("average sharpe ratio for each entire epoche is {} with standard error {}".format(mean(sharpe_ratio_total),st.sem(sharpe_ratio_total)))
fig,ax= plt.subplots(nrows=2,ncols=1,sharex=True,sharey=True)
l1=ax[0].hist(sharpe_first_half,bins=20)
ax[0].set_title('Annualised Sharpe Ratio (t=0:1000)')
l2=ax[1].hist(sharpe_sec_half,bins=20)
ax[1].set_title('Annualised Sharpe Ratio (t=1001:2000)')
plt.show()
#plt.hist(numTrades,bins=20)
#plt.plot(terminal_EMA_SharpeRatio)
#plt.show()
actionHist=env.actionHistory
ts=[t/100 for t in ts]
cum_log_r=cumsum([log(1+ts[i]) for i in range(len(ts))])
cum_log_R=cumsum([log(1+(actionHist[i]*ts[i])) for i in range(len(ts))])
fix, axes = plt.subplots(3, sharex=True)
ln1=axes[0].plot(cum_log_r,label='Buy and Hold')
ln2=axes[0].plot(cum_log_R,label='Trading Agent')
lns=ln1+ln2
labs=[l.get_label() for l in lns]
axes[0].legend(lns,labs,loc='upper left')
axes[0].set_ylabel("Cumulative Log Returns")
ax[0].set_title("Artificial Series")
ln3=axes[1].plot(actionHist,'r',label='Trades')
axes[1].set_ylabel("F(t)")
axes[2].plot(learner.ema_sharpeRatio)
axes[2].set_ylabel("EMA Sharpe Ratio")
plt.show()
def buildMinimalMDLSTMNetwork():
N = RecurrentNetwork('simpleMdLstmNet')
i = LinearLayer(4, name='i')
h = MDLSTMLayer(1, peepholes=True, name='mdlstm')
o = LinearLayer(1, name='o')
N.addInputModule(i)
N.addModule(h)
N.addOutputModule(o)
N.addConnection(IdentityConnection(i, h, outSliceTo=4))
N.addRecurrentConnection(
IdentityConnection(h, h, outSliceFrom=4, inSliceFrom=1))
N.addConnection(IdentityConnection(h, o, inSliceTo=1))
N.sortModules()
return N
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
from pybrain.structure import RecurrentNetwork
from pybrain.structure import FullConnection, LinearLayer, LSTMLayer
from parsemusic import ds
import random
print ds
layerCount = 10
net = RecurrentNetwork()
net.addInputModule(LinearLayer(10, name='in'))
for x in range(layerCount):
net.addModule(LSTMLayer(20, name='hidden' + str(x)))
net.addOutputModule(LinearLayer(10, name='out'))
net.addConnection(FullConnection(net['in'], net['hidden1'], name='cIn'))
for x in range(layerCount - 1):
net.addConnection(
FullConnection(net[('hidden' + str(x))],
net['hidden' + str(x + 1)],
name=('c' + str(x + 1))))
net.addConnection(
FullConnection(net['hidden' + str(layerCount - 1)],
net['out'],
name='cOut'))
net.sortModules()
from pybrain.supervised import RPropMinusTrainer
trainer = RPropMinusTrainer(net, dataset=ds)
epochcount = 0
while True:
startingnote = random.choice(range(1, 17))
startingnote2 = random.choice(range(1, 17))
def buildElmanNetwork(hiddenSize):
net = RecurrentNetwork()
inLayer = LinearLayer(sampleSize())
hiddenLayer = SigmoidLayer(hiddenSize)
outLayer = SigmoidLayer(outputSize())
net.addInputModule(inLayer)
net.addModule(hiddenLayer)
net.addOutputModule(outLayer)
hiddenRecursive = IdentityConnection(hiddenLayer, hiddenLayer)
inToHidden = FullConnection(inLayer, hiddenLayer)
hiddenToOut = FullConnection(hiddenLayer, outLayer)
net.addRecurrentConnection(hiddenRecursive)
net.addConnection(inToHidden)
net.addConnection(hiddenToOut)
net.sortModules()
net.randomize()
return net
def equivalence_recurrent(self, builder):
_net = pybrainbridge._RecurrentNetwork()
builder(_net)
net = RecurrentNetwork()
builder(net)
super(TestNetworkEquivalence, self).equivalence_recurrent(net, _net)
__author__ = 'davidoregan' from pybrain.structure import FeedForwardNetwork from pybrain.structure import LinearLayer, SigmoidLayer from pybrain.structure import FullConnection from pybrain.structure import RecurrentNetwork #Specifiy networks n = FeedForwardNetwork() r = RecurrentNetwork() LinearLayer(2, name="Nigger") inLayer = LinearLayer(2) hiddenLayer = SigmoidLayer(3) outLayer = LinearLayer(1) n.addInputModule(inLayer) n.addModule(hiddenLayer) n.addOutputModule(outLayer) in_to_hidden = FullConnection(inLayer, hiddenLayer) hidden_to_out = FullConnection(hiddenLayer, outLayer) n.addConnection(in_to_hidden) n.addConnection(hidden_to_out) n.sortModules() r.addInputModule(LinearLayer(2, name='in')) r.addModule(SigmoidLayer(3, name='hidden')) r.addOutputModule(LinearLayer(1, name='out'))
class RecurrentNeuralNetwork:
"""
Recurent neural network.
"""
def __init__(self, nin, nout):
singleton.append(self)
self.inn = nin
self.outn = nout
self.n = buildNetwork(nin, 20, nout, bias=False, recurrent=True)
self.n = RecurrentNetwork()
self.n.addInputModule(LinearLayer(nin, name='in'))
self.n.addOutputModule(LinearLayer(nout, name='out'))
self.n.addModule(SigmoidLayer(8, name='hidden2'))
self.n.addModule(TanhLayer(nin+nout/2, name='hidden1'))
self.n.addModule(BiasUnit(name='bias'))
self.n.addModule(LSTMLayer(5, name='memory'))
self.n.addConnection(FullConnection(self.n['in'], self.n['hidden1']))
self.n.addConnection(FullConnection(self.n['bias'], self.n['hidden1']))
self.n.addConnection(FullConnection(self.n['hidden1'], self.n['hidden2']))
self.n.addConnection(FullConnection(self.n['hidden2'], self.n['out']))
self.n.addConnection(FullConnection(self.n['hidden1'], self.n['memory']))
self.n.addConnection(FullConnection(self.n['memory'], self.n['hidden2']))
self.n.addConnection(FullConnection(self.n['in'], self.n['hidden2']))
self.n.addConnection(FullConnection(self.n['hidden2'], self.n['out']))
self.n.addRecurrentConnection(FullConnection(self.n['hidden1'], self.n['hidden1']))
self.n.addRecurrentConnection(FullConnection(self.n['memory'], self.n['hidden1']))
self.n.sortModules()
def set_learning_data(self, dataset):
"""
Set dataset used to train network.
"""
self.ds_learn = dataset
def train(self, epochs=100):
"""
Train the network
"""
#self.n.reset()
trainer = BackpropTrainer(self.n, self.ds_learn, verbose=True)
# trainer.setData(self.ds_learn)
return trainer.trainEpochs(epochs=epochs)
def validate_error(self, dataset):
"""
Return error value for given dataset
"""
v = Validator()
#self.n.reset()
return v.MSE(self.n, dataset)
def calculate(self, dataset):
"""
Return network response for given dataset
"""
#self.n.reset()
return self.n.activateOnDataset(dataset)
#testdf = pd.read_pickle(testdf_path)
# Construct one hot character dictionary
conglomerateString = ''
for index, row in df.iterrows():
conglomerateString += row.values
conglomerateSet = list(set(list(conglomerateString[0])))
codeTable = pd.Series(data=conglomerateSet, index=conglomerateSet)
codeTable = pd.get_dummies(codeTable)
conglomerateSet = []
conglomerateString = []
# Construct LSTM network
rnn = RecurrentNetwork()
inputSize = len(codeTable['a'].values)
outputSize = 4
hiddenSize = 10
rnn.addInputModule(LinearLayer(dim=inputSize, name='in'))
rnn.addModule(TanhLayer(dim=hiddenSize, name='in_proc'))
rnn.addModule(LSTMLayer(dim=hiddenSize, peepholes=True, name='hidden'))
rnn.addModule(TanhLayer(dim=hiddenSize, name='out_proc'))
rnn.addOutputModule(SoftmaxLayer(dim=outputSize, name='out'))
rnn.addConnection(FullConnection(rnn['in'], rnn['in_proc'], name='c1'))
rnn.addConnection(FullConnection(rnn['in_proc'], rnn['hidden'], name='c2'))
rnn.addRecurrentConnection(
FullConnection(rnn['hidden'], rnn['hidden'], name='c3'))
def buildMinimalLSTMNetwork():
N = RecurrentNetwork('simpleLstmNet')
i = LinearLayer(4, name='i')
h = LSTMLayer(1, peepholes=True, name='lstm')
o = LinearLayer(1, name='o')
N.addInputModule(i)
N.addModule(h)
N.addOutputModule(o)
N.addConnection(IdentityConnection(i, h))
N.addConnection(IdentityConnection(h, o))
N.sortModules()
return N
def construct_network(input_len, output_len, hidden_nodes, is_elman=True):
n = RecurrentNetwork()
n.addInputModule(LinearLayer(input_len, name="i"))
n.addModule(BiasUnit("b"))
n.addModule(SigmoidLayer(hidden_nodes, name="h"))
n.addOutputModule(LinearLayer(output_len, name="o"))
n.addConnection(FullConnection(n["i"], n["h"]))
n.addConnection(FullConnection(n["b"], n["h"]))
n.addConnection(FullConnection(n["b"], n["o"]))
n.addConnection(FullConnection(n["h"], n["o"]))
if is_elman:
# Elman (hidden->hidden)
n.addRecurrentConnection(FullConnection(n["h"], n["h"]))
else:
# Jordan (out->hidden)
n.addRecurrentConnection(FullConnection(n["o"], n["h"]))
n.sortModules()
n.reset()
return n
def trainNetwork(dirname):
numFeatures = 2000
ds = SequentialDataSet(numFeatures, 1)
tracks = glob.glob(os.path.join(dirname, "*.csv"))
for t in tracks:
track = os.path.splitext(t)[0]
# load training data
print "Reading %s..." % t
data = numpy.genfromtxt(t, delimiter=",")
numData = data.shape[0]
# add the input to the dataset
print "Adding to dataset..."
ds.newSequence()
for i in range(numData):
# ds.addSample(data[i], (labels[i],))
input = data[i]
label = input[numFeatures]
if label > 0:
label = midi_util.frequencyToMidi(label)
ds.addSample(input[0:numFeatures], (label,))
# initialize the neural network
print "Initializing neural network..."
# net = buildNetwork(numFeatures, 50, 1,
# hiddenclass=LSTMLayer, bias=True, recurrent=True)
# manual network building
net = RecurrentNetwork()
inlayer = LinearLayer(numFeatures)
# h1 = LSTMLayer(70)
# h2 = SigmoidLayer(50)
octaveLayer = LSTMLayer(5)
noteLayer = LSTMLayer(12)
combinedLayer = SigmoidLayer(60)
outlayer = LinearLayer(1)
net.addInputModule(inlayer)
net.addOutputModule(outlayer)
# net.addModule(h1)
# net.addModule(h2)
net.addModule(octaveLayer)
net.addModule(noteLayer)
net.addModule(combinedLayer)
# net.addConnection(FullConnection(inlayer, h1))
# net.addConnection(FullConnection(h1, h2))
# net.addConnection(FullConnection(h2, outlayer))
net.addConnection(FullConnection(inlayer, octaveLayer))
net.addConnection(FullConnection(inlayer, noteLayer))
# net.addConnection(FullConnection(octaveLayer,combinedLayer))
for i in range(5):
net.addConnection(
FullConnection(
octaveLayer, combinedLayer, inSliceFrom=i, inSliceTo=i + 1, outSliceFrom=i * 12, outSliceTo=(i + 1) * 12
)
)
net.addConnection(FullConnection(noteLayer, combinedLayer))
net.addConnection(FullConnection(combinedLayer, outlayer))
net.sortModules()
# train the network on the dataset
print "Training neural net"
trainer = RPropMinusTrainer(net, dataset=ds)
## trainer.trainUntilConvergence(maxEpochs=50, verbose=True, validationProportion=0.1)
error = -1
for i in range(150):
new_error = trainer.train()
print "error: " + str(new_error)
if abs(error - new_error) < 0.005:
break
error = new_error
# save the network
print "Saving neural network..."
NetworkWriter.writeToFile(net, os.path.basename(dirname) + "designnet")
def buildNetwork(N):
dimension = WINDOW_SIZE
inLayer = LinearLayer(dimension)
hiddenLayer = SigmoidLayer(N)
outLayer = LinearLayer(dimension)
# bias disabled, too much over training
#bias = BiasUnit(name='bias')
in_to_hidden = FullConnection(inLayer, hiddenLayer)
hidden_to_out = FullConnection(hiddenLayer, outLayer)
#bias_to_out = FullConnection(bias, outLayer)
#bias_to_hidden = FullConnection(bias, hiddenLayer)
net = RecurrentNetwork()
#net.addModule(bias)
net.addInputModule(inLayer)
net.addModule(hiddenLayer)
net.addOutputModule(outLayer)
net.addConnection(in_to_hidden)
net.addConnection(hidden_to_out)
net.addRecurrentConnection(FullConnection(hiddenLayer, hiddenLayer))
#net.addConnection(bias_to_hidden)
#net.addConnection(bias_to_out)
net.sortModules()
return net
def runNeuralLearningCurveSimulation(dataTrain, dataTest, train_tfidf,
test_tfidf, outFile):
print 'running neural learning curve'
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 = 200
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 buildMinimalLSTMNetwork():
N = RecurrentNetwork('simpleLstmNet')
i = LinearLayer(4, name='i')
h = LSTMLayer(1, peepholes=True, name='lstm')
o = LinearLayer(1, name='o')
N.addInputModule(i)
N.addModule(h)
N.addOutputModule(o)
N.addConnection(IdentityConnection(i, h))
N.addConnection(IdentityConnection(h, o))
N.sortModules()
return N
n.addOutputModule(outLayer) # Full Connection class - add connections/synapses in_to_hidden = FullConnection(inLayer, hiddenLayer) hidden_to_out = FullConnection(hiddenLayer, outLayer) n.addConnection(in_to_hidden) n.addConnection(hidden_to_out) #makes our MLP usable, n.sortModules() print n.activate([1, 2]) print n #Recureent Connection Class -which looks back in time one timestep. n = RecurrentNetwork() n.addInputModule(LinearLayer(2, name='in')) n.addModule(SigmoidLayer(3, name='hidden')) n.addOutputModule(LinearLayer(1, name='out')) n.addConnection(FullConnection(n['in'], n['hidden'], name='c1')) n.addConnection(FullConnection(n['hidden'], n['out'], name='c2')) n.sortModules() print n.activate((2, 2)) print n.activate((2, 2)) print n.activate((2, 2)) n.reset() print n.activate((2, 2)) ####################################### ######### Classification with feed forward networks
all weights of the network at once. """ print(hidden2out.params) print(n.params) """ The former are the last slice of the latter. """ print(n.params[-3:] == hidden2out.params) """ Ok, after having covered the basics, let's move on to some additional concepts. First of all, we encourage you to name all modules, or connections you create, because that gives you more readable printouts, and a very concise way of accessing them. We now build an equivalent network to the one before, but with a more concise syntax: """ n2 = RecurrentNetwork(name='net2') n2.addInputModule(LinearLayer(2, name='in')) n2.addModule(SigmoidLayer(3, name='h')) n2.addOutputModule(LinearLayer(1, name='out')) n2.addConnection(FullConnection(n2['in'], n2['h'], name='c1')) n2.addConnection(FullConnection(n2['h'], n2['out'], name='c2')) n2.sortModules() """ Printouts look more concise and readable: """ print(n2) """ There is an even quicker way to build networks though, as long as their structure is nothing more fancy than a stack of fully connected layers: """ n3 = buildNetwork(2, 3, 1, bias=False)
from pybrain.structure import FeedForwardNetwork, RecurrentNetwork, LinearLayer, SigmoidLayer, FullConnection #### #RECURRENT NETWORK #### n = RecurrentNetwork() inLayer = LinearLayer(inputVector, name='inputLayer') hiddenLayerA = SigmoidLayer(hiddenVector, name='hiddenLayerA') hiddenLayerB = SigmoidLayer(hiddenVector, name='hiddenLayerB') outputLayer = LinearLayer(outputVector, name='outputLayer') n.addInputModule(inLayer) n.addModule(hiddenLayerA) n.addModule(hiddenLayerB) n.addOutputModule(outputLayer) n.addConnection(FullConnection(n['inputLayer'], n['hiddenLayerA'], name='c1')) n.addConnection(FullConnection(n['hiddenLayerA'], n['hiddenLayerB'], name='c2')) n.addConnection(FullConnection(n['hiddenLayerB'], n['outputLayer'], name='c3')) n.addRecurrentConnection(FullConnection(n['hiddenLayerA'], n['hiddenLayerB'], name='rec3')) n.sortModules() print 'Network One (Recurrent)' + str(n.activate([1,2,3])) print 'Network One (Recurrent)' + str(n.activate([1,2,3])) #### #FEED FORWARD NETWORK ####
def exec_algo(xml_file, output_location):
rootObj=ml.parse(xml_file)
file_name=rootObj.MachineLearning.prediction.datafile
file=open(file_name)
var_input=rootObj.MachineLearning.prediction.input
var_output=rootObj.MachineLearning.prediction.output
var_classes=rootObj.MachineLearning.prediction.classes
DS=ClassificationDataSet(var_input,var_output,nb_classes=var_classes)
#DS1=ClassificationDataSet(13,1,nb_classes=10)
for line in file.readlines():
data=[float(x) for x in line.strip().split(',') if x != '']
inp=tuple(data[:var_input])
output=tuple(data[var_input:])
DS.addSample(inp,output)
tstdata,trndata=DS.splitWithProportion(0)
#trndatatest,tstdatatest=DS1.splitWithProportion(0)
trdata=ClassificationDataSet(trndata.indim,1,nb_classes=10)
#tsdata=ClassificationDataSet(DS1.indim,1,nb_classes=10)
#tsdata1=ClassificationDataSet(DS1.indim,1,nb_classes=10)
for i in xrange(trndata.getLength()):
if (trndata.getSample(i)[1][0]!=100):
trdata.addSample(trndata.getSample(i)[0],trndata.getSample(i)[1])
trdata._convertToOneOfMany()
#tsdata._convertToOneOfMany()
#tsdata1._convertToOneOfMany()
print "%d" % (trdata.getLength())
rnn=RecurrentNetwork()
inputLayer=LinearLayer(trdata.indim)
hiddenLayer=rootObj.MachineLearning.prediction.algorithm.RecurrentNeuralNetwork.hiddenLayerActivation
hiddenNeurons=rootObj.MachineLearning.prediction.algorithm.RecurrentNeuralNetwork.hiddenNeurons
if hiddenLayer=='Sigmoid':
hiddenLayer=SigmoidLayer(hiddenNeurons)
elif hiddenLayer=='Softmax':
hiddenLayer=SoftmaxLayer(hiddenNeurons)
else:
hiddenLayer=LinearLayer(hiddenNeurons)
outputLayer=rootObj.MachineLearning.prediction.algorithm.RecurrentNeuralNetwork.outputLayerActivation
if outputLayer=='Sigmoid':
outputLayer=SigmoidLayer(trdata.outdim)
elif outputLayer=='Softmax':
outputLayer=SoftmaxLayer(trdata.outdim)
else:
outputLayer=LinearLayer(trdata.outdim)
rnn.addInputModule(inputLayer)
rnn.addModule(hiddenLayer)
rnn.addOutputModule(outputLayer)
rnn_type=rootObj.MachineLearning.prediction.algorithm.RecurrentNeuralNetwork.RNN_Type
in_to_hidden=FullConnection(inputLayer,hiddenLayer)
hidden_to_outputLayer=FullConnection(hiddenLayer,outputLayer)
rnn.addConnection(in_to_hidden)
rnn.addConnection(hidden_to_outputLayer)
if rnn_type=='Elman':
hidden_to_hidden=FullConnection(hiddenLayer,hiddenLayer, name='c3')
rnn.addRecurrentConnection(hidden_to_hidden)
#hidden_to_hidden=FullConnection(hiddenLayer,hiddenLayer, name='c3')
if rnn_type=='Jordan':
output_to_hidden=FullConnection(outputLayer,hiddenLayer, name='c3')
rnn.addRecurrentConnection(output_to_hidden)
#rnn.addRecurrentConnection(hidden_to_hidden)
momentum=rootObj.MachineLearning.prediction.algorithm.RecurrentNeuralNetwork.momentum
weightdecay=rootObj.MachineLearning.prediction.algorithm.RecurrentNeuralNetwork.learningRate
rnn.sortModules()
trainer=BackpropTrainer(rnn,dataset=trdata,momentum=0.1,verbose=True,weightdecay=0.01)
trainer.train();
result=(percentError(trainer.testOnClassData(dataset=trdata),trdata['class']))
#result1=percentError(trainer.testOnClassData(dataset=tsdata1),tsdata1['class'])
print ('%f \n') % (100-result)
#print ('%f \n') % (100-result1)
ts=time.time()
directory = output_location + sep + str(int(ts))
makedirs(directory)
fileObject=open(output_location + sep + str(int(ts)) + sep + 'pybrain_RNN','w')
pickle.dump(trainer,fileObject)
pickle.dump(rnn,fileObject)
fileObject.close()
#inLayer = LinearLayer(ds.indim) #hiddenLayer = SigmoidLayer(5) #outLayer = SoftmaxLayer(ds.outdim) #net.addInputModule(inLayer) #net.addModule(hiddenLayer) #net.addOutputModule(outLayer) #from pybrain.structure import FullConnection #in_to_hidden = FullConnection(inLayer, hiddenLayer) #hidden_to_out = FullConnection(hiddenLayer, outLayer) #net.addConnection(in_to_hidden) #net.addConnection(hidden_to_out) #net.sortModules() #net = buildNetwork(ds.indim, 2, ds.outdim, outclass=SoftmaxLayer) from pybrain.structure import RecurrentNetwork net = RecurrentNetwork() net.addInputModule(LinearLayer(ds.indim, name='inLayer')) net.addModule(SigmoidLayer(ds.indim, name='hiddenLayer')) net.addOutputModule(SoftmaxLayer(ds.outdim, name='outLayer')) net.addConnection(FullConnection(net['inLayer'], net['hiddenLayer'], name='in_to_hidden')) net.addConnection(FullConnection(net['hiddenLayer'], net['outLayer'], name='hidden_to_out')) net.addRecurrentConnection(FullConnection(net['hiddenLayer'], net['hiddenLayer'], name='hidden_to_hidden')) net.sortModules() #Train net from pybrain.supervised.trainers import BackpropTrainer trainer = BackpropTrainer(net, ds, momentum=0.1, verbose=True, weightdecay=0.01) #for i in range(10): # if i%20==0: # print i
def buildMixedNestedNetwork():
""" build a nested network with the inner one being a ffn and the outer one being recurrent. """
N = RecurrentNetwork('outer')
a = LinearLayer(1, name = 'a')
b = LinearLayer(2, name = 'b')
c = buildNetwork(2, 3, 1)
c.name = 'inner'
N.addInputModule(a)
N.addModule(c)
N.addOutputModule(b)
N.addConnection(FullConnection(a,b))
N.addConnection(FullConnection(b,c))
N.addRecurrentConnection(FullConnection(c,c))
N.sortModules()
return N
class LanguageLearner:
__OUTPUT = "Sample at {0} epochs (prompt=\"{1}\", length={2}): {3}"
def __init__(self, trainingText, hiddenLayers, hiddenNodes):
self.__initialized = False
with open(trainingText) as f:
self.raw = f.read()
self.characters = list(self.raw)
self.rawData = list(map(ord, self.characters))
print("Creating alphabet mapping...")
self.mapping = []
for charCode in self.rawData:
if charCode not in self.mapping:
self.mapping.append(charCode)
print("Mapping of " + str(len(self.mapping)) + " created.")
print(str(self.mapping))
print("Converting data to mapping...")
self.data = []
for charCode in self.rawData:
self.data.append(self.mapping.index(charCode))
print("Done.")
self.dataIn = self.data[:-1:]
self.dataOut = self.data[1::]
self.inputs = 1
self.hiddenLayers = hiddenLayers
self.hiddenNodes = hiddenNodes
self.outputs = 1
def initialize(self, verbose):
print("Initializing language learner...")
self.verbose = verbose
# Create network and modules
self.net = RecurrentNetwork()
inp = LinearLayer(self.inputs, name="in")
hiddenModules = []
for i in range(0, self.hiddenLayers):
hiddenModules.append(LSTMLayer(self.hiddenNodes, name=("hidden-" + str(i + 1))))
outp = LinearLayer(self.outputs, name="out")
# Add modules to the network with recurrence
self.net.addOutputModule(outp)
self.net.addInputModule(inp)
for module in hiddenModules:
self.net.addModule(module)
# Create connections
self.net.addConnection(FullConnection(self.net["in"], self.net["hidden-1"]))
for i in range(0, len(hiddenModules) - 1):
self.net.addConnection(FullConnection(self.net["hidden-" + str(i + 1)], self.net["hidden-" + str(i + 2)]))
self.net.addRecurrentConnection(FullConnection(self.net["hidden-" + str(i + 1)], self.net["hidden-" + str(i + 1)]))
self.net.addRecurrentConnection(FullConnection(self.net["hidden-" + str(len(hiddenModules))],
self.net["hidden-" + str(len(hiddenModules))]))
self.net.addConnection(FullConnection(self.net["hidden-" + str(len(hiddenModules))], self.net["out"]))
self.net.sortModules()
self.trainingSet = SequentialDataSet(self.inputs, self.outputs)
for x, y in zip(self.dataIn, self.dataOut):
self.trainingSet.newSequence()
self.trainingSet.appendLinked([x], [y])
self.net.randomize()
print("Neural network initialzed with structure:")
print(self.net)
self.trainer = BackpropTrainer(self.net, self.trainingSet, verbose=verbose)
self.__initialized = True
print("Successfully initialized network.")
def train(self, epochs, frequency, prompt, length):
if not self.__initialized:
raise Exception("Attempted to train uninitialized LanguageLearner")
print ("Beginning training for " + str(epochs) + " epochs...")
if frequency >= 0:
print(LanguageLearner.__OUTPUT.format(0, prompt, length, self.sample(prompt, length)))
for i in range(1, epochs):
print("Error at " + str(i) + " epochs: " + str(self.trainer.train()))
if i % frequency == 0:
print(LanguageLearner.__OUTPUT.format(i, prompt, length, self.sample(prompt, length)))
print("Completed training.")
def sample(self, prompt, length):
self.net.reset()
if prompt == None:
prompt = chr(random.choice(self.mapping))
output = prompt
charCode = ord(prompt)
for i in range(0, length):
sampledResult = self.net.activate([charCode])
charCode = int(round(sampledResult[0]))
if charCode < 0 or charCode >= len(self.mapping):
return output + "#TERMINATED_SAMPLE(reason: learner guessed invalid character)"
output += chr(self.mapping[charCode])
return output
def trainedRNN():
n = RecurrentNetwork()
n.addInputModule(LinearLayer(4, name='in'))
n.addModule(SigmoidLayer(6, name='hidden'))
n.addOutputModule(LinearLayer(2, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(NMConnection(n['out'], n['out'], name='nmc'))
# n.addRecurrentConnection(FullConnection(n['out'], n['hidden'], inSliceFrom = 0, inSliceTo = 1, outSliceFrom = 0, outSliceTo = 3))
n.sortModules()
draw_connections(n)
d = getDatasetFromFile(root.path() + "/res/dataSet")
t = BackpropTrainer(n, d, learningrate=0.001, momentum=0.75)
t.trainOnDataset(d)
count = 0
while True:
globErr = t.train()
print globErr
if globErr < 0.01:
break
count += 1
if count == 50:
return trainedRNN()
# exportRNN(n)
draw_connections(n)
return n
intuple.append(int(item)) for item in sentence[0]: if int(item) < 1: outtuple.append(-1) else: outtuple.append(int(item)) print intuple, outtuple result = rnet.activate(intuple) if (result != sentence[0]) print "FAIL:",sentence exit(0) #basically main if __name__ == "__main__": #Initialize the ANN rnet = RecurrentNetwork() rnet.addInputModule(LinearLayer(8, name="word_hash")) rnet.addInputModule(LinearLayer(6, name="word_type")) rnet.addModule(LSTMLayer(6, name="hidden")) rnet.addOutputModule(TanhLayer(6, name="out")) rnet.addConnection(FullConnection(rnet["word_type"], rnet["hidden"])) rnet.addRecurrentConnection(FullConnection(rnet["hidden"], rnet["hidden"])) rnet.addConnection(FullConnection(rnet["hidden"], rnet["out"])) rnet.addConnection(FullConnection(rnet["word_hash"], rnet["out"])) rnet.sortModules() #Initialize the dataset sds = SequenceClassificationDataSet(14,6)
def trained_cat_dog_RFCNN():
n = RecurrentNetwork()
d = get_cat_dog_trainset()
input_size = d.getDimension('input')
n.addInputModule(LinearLayer(input_size, name='in'))
n.addModule(SigmoidLayer(input_size + 1500, name='hidden'))
n.addOutputModule(LinearLayer(2, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(FullConnection(n['out'], n['hidden'], name='nmc'))
n.sortModules()
t = BackpropTrainer(n, d, learningrate=0.0001) #, momentum=0.75)
count = 0
while True:
globErr = t.train()
print globErr
count += 1
if globErr < 0.01:
break
if count == 30:
break
exportCatDogRFCNN(n)
return n
from pybrain.datasets import SupervisedDataSet
from pybrain.supervised.trainers import BackpropTrainer
duration = 1
num_coeff = 26
max_freq = 8000
min_freq = 0
melArray = np.linspace(FEXT.freqToMel(min_freq), FEXT.freqToMel(max_freq),
num_coeff + 2)
ferqArray = FEXT.melToFreq(melArray)
freqArray_bin = np.floor(513 * ferqArray / 16000)
centralPoints = freqArray_bin[1:21]
freqbank = np.zeros((26, 257))
LSTMre = RecurrentNetwork()
LSTMre.addInputModule(LinearLayer(39, name='input'))
LSTMre.addModule(LSTMLayer(50, name='LSTM_hidden'))
LSTMre.addOutputModule(SoftmaxLayer(5, name='out'))
LSTMre.addConnection(
FullConnection(LSTMre['input'], LSTMre['LSTM_hidden'], name='c1'))
LSTMre.addConnection(
FullConnection(LSTMre['LSTM_hidden'], LSTMre['out'], name='c2'))
LSTMre.addRecurrentConnection(
FullConnection(LSTMre['LSTM_hidden'], LSTMre['LSTM_hidden'], name='c3'))
LSTMre.sortModules()
ds = SupervisedDataSet(39, 5)
#ser.
def trainedRNN():
n = RecurrentNetwork()
n.addInputModule(LinearLayer(4, name='in'))
n.addModule(SigmoidLayer(6, name='hidden'))
n.addOutputModule(LinearLayer(2, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(NMConnection(n['out'], n['out'], name='nmc'))
# n.addRecurrentConnection(FullConnection(n['out'], n['hidden'], inSliceFrom = 0, inSliceTo = 1, outSliceFrom = 0, outSliceTo = 3))
n.sortModules()
draw_connections(n)
d = getDatasetFromFile(root.path()+"/res/dataSet")
t = BackpropTrainer(n, d, learningrate=0.001, momentum=0.75)
t.trainOnDataset(d)
count = 0
while True:
globErr = t.train()
print globErr
if globErr < 0.01:
break
count += 1
if count == 50:
return trainedRNN()
# exportRNN(n)
draw_connections(n)
return n
def buildParityNet():
net = RecurrentNetwork()
net.addInputModule(LinearLayer(1, name = 'i'))
net.addModule(TanhLayer(2, name = 'h'))
net.addModule(BiasUnit('bias'))
net.addOutputModule(TanhLayer(1, name = 'o'))
net.addConnection(FullConnection(net['i'], net['h']))
net.addConnection(FullConnection(net['bias'], net['h']))
net.addConnection(FullConnection(net['bias'], net['o']))
net.addConnection(FullConnection(net['h'], net['o']))
net.addRecurrentConnection(FullConnection(net['o'], net['h']))
net.sortModules()
p = net.params
p[:] = [-0.5, -1.5, 1, 1, -1, 1, 1, -1, 1]
p *= 10.
return net
def trained_cat_dog_RFCNN():
n = RecurrentNetwork()
d = get_cat_dog_trainset()
input_size = d.getDimension('input')
n.addInputModule(LinearLayer(input_size, name='in'))
n.addModule(SigmoidLayer(input_size+1500, name='hidden'))
n.addOutputModule(LinearLayer(2, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.addRecurrentConnection(FullConnection(n['out'], n['hidden'], name='nmc'))
n.sortModules()
t = BackpropTrainer(n, d, learningrate=0.0001)#, momentum=0.75)
count = 0
while True:
globErr = t.train()
print globErr
count += 1
if globErr < 0.01:
break
if count == 30:
break
exportCatDogRFCNN(n)
return n
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)
all weights of the network at once. """ print hidden2out.params print n.params """ The former are the last slice of the latter. """ print n.params[-3:] == hidden2out.params """ Ok, after having covered the basics, let's move on to some additional concepts. First of all, we encourage you to name all modules, or connections you create, because that gives you more readable printouts, and a very concise way of accessing them. We now build an equivalent network to the one before, but with a more concise syntax: """ n2 = RecurrentNetwork(name='net2') n2.addInputModule(LinearLayer(2, name='in')) n2.addModule(SigmoidLayer(3, name='h')) n2.addOutputModule(LinearLayer(1, name='out')) n2.addConnection(FullConnection(n2['in'], n2['h'], name='c1')) n2.addConnection(FullConnection(n2['h'], n2['out'], name='c2')) n2.sortModules() """ Printouts look more concise and readable: """ print n2 """ There is an even quicker way to build networks though, as long as their structure is nothing more fancy than a stack of fully connected layers: """ n3 = buildNetwork(2, 3, 1, bias=False)
def buildSimpleLSTMNetwork(peepholes = False):
N = RecurrentNetwork('simpleLstmNet')
i = LinearLayer(100, name = 'i')
h = LSTMLayer(10, peepholes = peepholes, name = 'lstm')
o = LinearLayer(1, name = 'o')
b = BiasUnit('bias')
N.addModule(b)
N.addOutputModule(o)
N.addInputModule(i)
N.addModule(h)
N.addConnection(FullConnection(i, h, name = 'f1'))
N.addConnection(FullConnection(b, h, name = 'f2'))
N.addRecurrentConnection(FullConnection(h, h, name = 'r1'))
N.addConnection(FullConnection(h, o, name = 'r1'))
N.sortModules()
return N
from pybrain.rl.environments.timeseries.maximizereturntask import MaximizeReturnTask from pybrain.rl.environments.timeseries.timeseries import MonthlySnPEnvironment from pybrain.rl.learners.directsearch.rrl import RRL from pybrain.structure import RecurrentNetwork from pybrain.structure import LinearLayer, SigmoidLayer, TanhLayer, BiasUnit from pybrain.structure import FullConnection from pybrain.rl.agents import LearningAgent from pybrain.rl.experiments import EpisodicExperiment from numpy import sign, round from matplotlib import pyplot net= RecurrentNetwork() #Single linear layer with bias unit, and single tanh layer. the linear layer is whats optimised net.addInputModule(BiasUnit(name='bias')) net.addOutputModule(TanhLayer(1, name='out')) net.addRecurrentConnection(FullConnection(net['out'], net['out'], name='c3')) net.addInputModule(LinearLayer(1,name='in')) net.addConnection(FullConnection(net['in'],net['out'],name='c1')) net.addConnection((FullConnection(net['bias'],net['out'],name='c2'))) net.sortModules() net._setParameters([-8.79227886e-02, -8.29319017e+02, 1.25946474e+00]) print(net._params) env=MonthlySnPEnvironment() task=MaximizeReturnTask(env) learner = RRL() # ENAC() #Q_LinFA(2,1) agent = LearningAgent(net,learner) exp=EpisodicExperiment(task,agent) exp.doEpisodes(10)
def buildNonGravityNet(recurrent = False):
if recurrent:
net = RecurrentNetwork()
else:
net = FeedForwardNetwork()
l1 = LinearLayer(2)
l2 = LinearLayer(3)
s1 = SigmoidLayer(2)
l3 = LinearLayer(1)
net.addInputModule(l1)
net.addModule(l2)
net.addModule(s1)
net.addOutputModule(l3)
net.addConnection(IdentityConnection(l1, l2, outSliceFrom = 1))
net.addConnection(IdentityConnection(l1, l2, outSliceTo = 2))
net.addConnection(IdentityConnection(l2, l3, inSliceFrom = 2))
net.addConnection(IdentityConnection(l2, l3, inSliceTo = 1))
net.addConnection(IdentityConnection(l1, s1))
net.addConnection(IdentityConnection(l2, s1, inSliceFrom = 1))
net.addConnection(IdentityConnection(s1, l3, inSliceFrom = 1))
if recurrent:
net.addRecurrentConnection(IdentityConnection(s1, l1))
net.addRecurrentConnection(IdentityConnection(l2, l2, inSliceFrom = 1, outSliceTo = 2))
net.sortModules()
return net
Xmax=np.asarray([x*1.0 for x in np.amax(X,axis=0)])
#X = (X-Xmin)/(Xmax-Xmin)
X=X/Xmax
Ymin=np.asarray([x*1.0 for x in np.amin(Y,axis=0)])
Ymax=np.asarray([x*1.0 for x in np.amax(Y,axis=0)])
Y = (Y-Ymin)/(Ymax-Ymin)
from pybrain.datasets import SupervisedDataSet
DS = SupervisedDataSet(4, 1)
for i in range(0, Y.size):
DS.addSample((X[i][0], X[i][1], X[i][2], X[i][3]), (float(Y[i]),))
## ----------------------- ANN ---------------------------- ##
from pybrain.structure import RecurrentNetwork
n = RecurrentNetwork()
from pybrain.structure import LinearLayer, SigmoidLayer
from pybrain.structure import FullConnection
n.addInputModule(SigmoidLayer(4, name='in'))
n.addModule(SigmoidLayer(3, name='hidden'))
n.addOutputModule(LinearLayer(1, name='out'))
n.addConnection(FullConnection(n['in'], n['hidden'], name='c1'))
n.addConnection(FullConnection(n['hidden'], n['out'], name='c2'))
n.sortModules() #initialisation
## ----------------------- Trainer ---------------------------- ##
def testTraining():
# the AnBnCn dataset (sequential)
d = AnBnCnDataSet()
# build a recurrent network to be trained
hsize = 2
n = RecurrentNetwork()
n.addModule(TanhLayer(hsize, name='h'))
n.addModule(BiasUnit(name='bias'))
n.addOutputModule(LinearLayer(1, name='out'))
n.addConnection(FullConnection(n['bias'], n['h']))
n.addConnection(FullConnection(n['h'], n['out']))
n.addRecurrentConnection(FullConnection(n['h'], n['h']))
n.sortModules()
# initialize the backprop trainer and train
t = BackpropTrainer(n, learningrate=0.1, momentum=0.0, verbose=True)
t.trainOnDataset(d, 200)
# the resulting weights are in the network:
print('Final weights:', n.params)
