def train(self, params):
"""
Train LSTM network on buffered dataset history
After training, run LSTM on history[:-1] to get the state correct
:param params:
:return:
"""
if params['reset_every_training']:
n = params['encoding_num']
self.net = buildNetwork(n,
params['num_cells'],
n,
hiddenclass=LSTMLayer,
bias=True,
outputbias=params['output_bias'],
recurrent=True)
self.net.reset()
# prepare training dataset
ds = SequentialDataSet(params['encoding_num'], params['encoding_num'])
history = self.window(self.history, params)
resets = self.window(self.resets, params)
for i in xrange(1, len(history)):
if not resets[i - 1]:
ds.addSample(self.encoder.encode(history[i - 1]),
self.encoder.encode(history[i]))
if resets[i]:
ds.newSequence()
if params['num_epochs'] > 1:
trainer = RPropMinusTrainer(self.net,
dataset=ds,
verbose=params['verbosity'] > 0)
if len(history) > 1:
trainer.trainEpochs(params['num_epochs'])
# run network on buffered dataset after training to get the state right
self.net.reset()
for i in xrange(len(history) - 1):
symbol = history[i]
output = self.net.activate(self.encoder.encode(symbol))
self.encoder.classify(output, num=params['num_predictions'])
if resets[i]:
self.net.reset()
else:
self.trainer.setData(ds)
self.trainer.train()
# run network on buffered dataset after training to get the state right
self.net.reset()
for i in xrange(len(history) - 1):
symbol = history[i]
output = self.net.activate(self.encoder.encode(symbol))
self.encoder.classify(output, num=params['num_predictions'])
if resets[i]:
self.net.reset()
def trainedLSTMNN2():
"""
n = RecurrentNetwork()
inp = LinearLayer(100, name = 'input')
hid = LSTMLayer(30, name='hidden')
out = LinearLayer(1, name='output')
#add modules
n.addOutputModule(out)
n.addInputModule(inp)
n.addModule(hid)
#add connections
n.addConnection(FullConnection(inp, hid))
n.addConnection(FullConnection(hid, out))
n.addRecurrentConnection(FullConnection(hid, hid))
n.sortModules()
"""
n = buildSimpleLSTMNetwork()
print "Network created"
d = load1OrderDataSet()
print "Data loaded"
t = RPropMinusTrainer(n, dataset=d, verbose=True)
t.trainUntilConvergence()
exportANN(n)
return n
def trainLSTMnet(net, numTrainSequence, seedSeq=1):
np.random.seed(seedSeq)
for _ in xrange(numTrainSequence):
(ds, in_seq, out_seq) = getReberDS(maxLength)
print("train seq", _, sequenceToWord(in_seq))
trainer = RPropMinusTrainer(net, dataset=ds)
trainer.trainEpochs(rptPerSeq)
return net
def train(self, ds, epochs_per_cycle, cycles):
trainer = RPropMinusTrainer(self.n, dataset=ds)
train_errors = []
for i in xrange(cycles):
trainer.trainEpochs(epochs_per_cycle)
train_errors.append(trainer.testOnData())
epoch = (i + 1) * epochs_per_cycle
print("\r epoch {}/{}".format(epoch, epochs_per_cycle * cycles))
sys.stdout.flush()
print("Final Error: " + str(train_errors[-1]))
return train_errors[-1]
def ltsmXY(tin, tout, title='ltsm.png'):
#datain = zip(tin[:-3], tin[1:-2], tin[2:-1])
#datain = zip(tin[:-8], tin[1:-7], tin[2:-6], tin[3:-5], tin[4:-4], tin[5:-3],tin[6:-2], tin[7:-1])
#datain = zip(tin[:-12], tin[1:-11], tin[2:-10], tin[3:-9], tin[4:-8], tin[5:-7],tin[6:-6], tin[7:-5], tin[8:-4], tin[9:-3], tin[10:-2], tin[11:-1])
datain = zip(tin[:-16], tin[1:-15], tin[2:-14], tin[3:-13], tin[4:-12], tin[5:-11],tin[6:-10], tin[7:-9], tin[8:-8], tin[9:-7], tin[10:-6], tin[11:-5], tin[12:-4], tin[13:-3], tin[14:-2], tin[15:-1])
#dataout = tout[3:]
#dataout = tout[8:]
#dataout = tout[12:]
dataout = tout[16:]
#ds = SequentialDataSet(3, 1)
#ds = SequentialDataSet(8, 1)
#ds = SequentialDataSet(12, 1)
ds = SequentialDataSet(16, 1)
for x, y in zip(datain[:len(datain)/2], dataout[:len(datain)/2]):
ds.addSample(x, y)
# add layers until overfitting the training data
#net = buildNetwork(3,5,1,hiddenclass=LSTMLayer, outputbias=False, recurrent=True)
#net = buildNetwork(8, 8, 1, hiddenclass=LSTMLayer, outputbias=False, recurrent=True)
#net = buildNetwork(12, 20, 1, hiddenclass=LSTMLayer, outputbias=False, recurrent=True)
net = buildNetwork(16, 20, 1, hiddenclass=LSTMLayer, outputbias=False, recurrent=True)
trainer = RPropMinusTrainer(net, dataset=ds)
train_errors = []
EPOCHS_PER_CYCLE = 5
CYCLES = 100
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
train_errors.append(trainer.testOnData())
epoch = (i+1) * EPOCHS_PER_CYCLE
#print "\r epoch {}/{}".format(epoch, EPOCHS)
stdout.flush()
print "final error =", train_errors[-1]
pred_out = []
for i in range(len(datain)):
pred_out.append(net.activate(datain[i]))
fig = plt.figure()
#tout[16:].plot(ax=ax, title='Occupancy')
plt.plot(tout[16:].index, tout[16:], 'y', linewidth=1.5)
plt.plot(tout[16:].index, pred_out, 'b+')
plt.legend(['Occupancy', 'LTSM'])
fig.tight_layout()
plt.savefig(title,inches='tight')
def __init__(self, indim, outdim):
# construct LSTM network - note the missing output bias
rnn = buildNetwork(indim,
5,
outdim,
hiddenclass=LSTMLayer,
outclass=SoftmaxLayer,
outputbias=False,
recurrent=True)
rnn2 = buildNetwork
# define a training method
trainer = RPropMinusTrainer(rnn)
def train(self, params, verbose=False):
if params['reset_every_training']:
if verbose:
print 'create lstm network'
random.seed(6)
if params['output_encoding'] == None:
self.net = buildNetwork(self.nDimInput,
params['num_cells'],
self.nDimOutput,
hiddenclass=LSTMLayer,
bias=True,
outputbias=True,
recurrent=True)
elif params['output_encoding'] == 'likelihood':
self.net = buildNetwork(self.nDimInput,
params['num_cells'],
self.nDimOutput,
hiddenclass=LSTMLayer,
bias=True,
outclass=SigmoidLayer,
recurrent=True)
self.net.reset()
ds = SequentialDataSet(self.nDimInput, self.nDimOutput)
networkInput = self.window(self.networkInput, params)
targetPrediction = self.window(self.targetPrediction, params)
# prepare a training data-set using the history
for i in xrange(len(networkInput)):
ds.addSample(self.inputEncoder.encode(networkInput[i]),
self.outputEncoder.encode(targetPrediction[i]))
if params['num_epochs'] > 1:
trainer = RPropMinusTrainer(self.net, dataset=ds, verbose=verbose)
if verbose:
print " train LSTM on ", len(
ds), " records for ", params['num_epochs'], " epochs "
if len(networkInput) > 1:
trainer.trainEpochs(params['num_epochs'])
else:
self.trainer.setData(ds)
self.trainer.train()
# run through the training dataset to get the lstm network state right
self.net.reset()
for i in xrange(len(networkInput)):
self.net.activate(ds.getSample(i)[0])
def ltsm(data):
from pybrain.datasets import SequentialDataSet
from itertools import cycle
datain = zip(data[:-6], data[1:-5], data[2:-4], data[3:-3], data[4:-2], data[5:-1])
dataout = data[6:]
ds = SequentialDataSet(6, 1)
for x, y in zip(datain, dataout):
ds.addSample(x, y)
from pybrain.tools.shortcuts import buildNetwork
from pybrain.structure.modules import LSTMLayer
net = buildNetwork(6, 7, 1, hiddenclass=LSTMLayer, outputbias=False, recurrent=True)
from pybrain.supervised import RPropMinusTrainer
from sys import stdout
trainer = RPropMinusTrainer(net, dataset=ds)
train_errors = []
EPOCHS_PER_CYCLE = 5
CYCLES = 100
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
train_errors.append(trainer.testOnData())
epoch = (i+1) * EPOCHS_PER_CYCLE
#print "\r epoch {}/{}".format(epoch, EPOCHS)
stdout.flush()
print "final error =", train_errors[-1]
'''
plt.figure()
plt.plot(range(0, EPOCHS, EPOCHS_PER_CYCLE), train_errors)
plt.xlabel('epoch')
plt.ylabel('error')
plt.show()
'''
test_error = 0.
cnt = 0
for sample, target in ds.getSequenceIterator(0):
#print "sample = ", sample
#print "predicted next sample = %4.1f" % net.activate(sample)
#print "actual next sample = %4.1f" % target
test_error += abs(net.activate(sample) - target)
cnt += 1
test_error /= cnt
print "test (train) error =", test_error
def handle(self, *args, **options):
ticker = args[0]
print("****** STARTING PREDICTOR " + ticker + " ******* ")
prices = Price.objects.filter(
symbol=ticker).order_by('-created_on').values_list('price',
flat=True)
data = normalization(list(prices[0:NUM_MINUTES_BACK].reverse()))
data = [int(x * MULT_FACTOR) for x in data]
print(data)
ds = SupervisedDataSet(5, 1)
try:
for i, val in enumerate(data):
DS.addSample((data[i], data[i + 1], data[i + 2], data[i + 3],
data[i + 4]), (data[i + 5], ))
except Exception:
pass
net = buildNetwork(5,
40,
1,
hiddenclass=LSTMLayer,
outputbias=False,
recurrent=True)
trainer = RPropMinusTrainer(net, dataset=ds)
train_errors = [] # save errors for plotting later
EPOCHS_PER_CYCLE = 5
CYCLES = 100
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
train_errors.append(trainer.testOnData())
epoch = (i + 1) * EPOCHS_PER_CYCLE
print("\r epoch {}/{}".format(epoch, EPOCHS), end="")
stdout.flush()
print()
print("final error =", train_errors[-1])
for sample, target in ds.getSequenceIterator(0):
show_pred_sample = net.activate(sample) / MULT_FACTOR
show_sample = sample / MULT_FACTOR
show_target = target / MULT_FACTOR
show_diff = show_pred_sample - show_target
show_diff_pct = 100 * show_diff / show_pred_sample
print("{} => {}, act {}. ({}%)".format(
show_sample[0], round(show_pred_sample[0], 3), show_target[0],
int(round(show_diff_pct[0], 0))))
def train(context, trainX, trainY):
ds = SequentialDataSet(4, 1)
for dataX, dataY in zip(trainX, trainY):
ds.addSample(dataX, dataY)
net = buildNetwork(4,
1,
1,
hiddenclass=LSTMLayer,
outputbias=False,
recurrent=True)
trainer = RPropMinusTrainer(net, dataset=ds)
EPOCHS_PER_CYCLE = 5
CYCLES = 5
for i in range(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
return net, trainer.testOnData()
def updateValue(self,laststate, state, lastaction,lastreward,lbda=1):
if self.t < self.horizon: return
qvalues = self.getValues()
if qvalues is None: return
qvalue = qvalues[lastaction]
next_qvalues = self.getTargetValues()
max_q_index = np.argmax(next_qvalues)
maxnext = next_qvalues[max_q_index]
if self.nn:
update = (lastreward + (self.gamma * maxnext))
qvalues[lastaction] = update
from pybrain.supervised import RPropMinusTrainer
trainer = RPropMinusTrainer(self.nn)
dataset = Sequential()
trainer.trainOnDataset(dataset)
else:
self.Q[laststate][lastaction]=qvalue + self.alpha * lbda * (lastreward + self.gamma * maxnext - qvalue)
def train(ds, net):
# Train the network
trainer = RPropMinusTrainer(net, dataset=ds)
train_errors = [] # save errors for plotting later
EPOCHS_PER_CYCLE = 5
CYCLES = 100
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
error = trainer.testOnData()
train_errors.append(error)
epoch = (i + 1) * EPOCHS_PER_CYCLE
print("\r epoch {}/{}".format(epoch, EPOCHS))
stdout.flush()
# print("final error =", train_errors[-1])
return train_errors, EPOCHS, EPOCHS_PER_CYCLE
def main():
generated_data = [0 for i in range(10000)]
rate, data = get_data_from_wav("../../data/natabhairavi_violin.wav")
data = data[1000:190000]
print("Got wav")
ds = SequentialDataSet(1, 1)
for sample, next_sample in zip(data, cycle(data[1:])):
ds.addSample(sample, next_sample)
net = buildNetwork(1,
5,
1,
hiddenclass=LSTMLayer,
outputbias=False,
recurrent=True)
trainer = RPropMinusTrainer(net, dataset=ds)
train_errors = [] # save errors for plotting later
EPOCHS_PER_CYCLE = 5
CYCLES = 10
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
train_errors.append(trainer.testOnData())
epoch = (i + 1) * EPOCHS_PER_CYCLE
print("\r epoch {}/{}".format(epoch, EPOCHS), end="")
stdout.flush()
# predict new values
old_sample = [100]
for i in xrange(500000):
new_sample = net.activate(old_sample)
old_sample = new_sample
generated_data[i] = new_sample[0]
print(new_sample)
wavfile.write("../../output/test.wav", rate, np.array(generated_data))
def train(self, params):
n = params['encoding_num']
net = buildNetwork(n,
params['num_cells'],
n,
hiddenclass=LSTMLayer,
bias=True,
outputbias=params['output_bias'],
recurrent=True)
net.reset()
ds = SequentialDataSet(n, n)
trainer = RPropMinusTrainer(net, dataset=ds)
history = self.window(self.history, params)
resets = self.window(self.resets, params)
for i in xrange(1, len(history)):
if not resets[i - 1]:
ds.addSample(self.encoder.encode(history[i - 1]),
self.encoder.encode(history[i]))
if resets[i]:
ds.newSequence()
if len(history) > 1:
trainer.trainEpochs(params['num_epochs'])
net.reset()
for i in xrange(len(history) - 1):
symbol = history[i]
output = net.activate(self.encoder.encode(symbol))
predictions = self.encoder.classify(output,
num=params['num_predictions'])
if resets[i]:
net.reset()
return net
def main():
config = MU.ConfigReader('configs/%s' % sys.argv[1])
config.read()
logDir = '%s-%s' % (__file__, sys.argv[1])
os.mkdir(logDir)
with open('%s/config.txt' % logDir, 'w') as outfile:
json.dump(config.getConfigDict(), outfile, indent=4)
dr = MU.DataReader(config['input_tsv_path'])
data = dr.read(config['interested_columns'])
inLabels = config['input_columns']
outLabels = config['output_columns']
tds, vds = seqDataSetPair(data, inLabels, outLabels,
config['seq_label_column'], config['test_seqno'],
config['validation_seqno'])
inScale = config.getDataScale(inLabels)
outScale = config.getDataScale(outLabels)
normalizeDataSet(tds, ins=inScale, outs=outScale)
normalizeDataSet(vds, ins=inScale, outs=outScale)
trainData = tds
validationData = vds
fdim = tds.indim / 5 + 5
xdim = tds.outdim * 2
rnn = buildNetwork(tds.indim,
fdim,
fdim,
fdim,
xdim,
tds.outdim,
hiddenclass=SigmoidLayer,
recurrent=True)
rnn.addRecurrentConnection(FullConnection(rnn['hidden0'], rnn['hidden0']))
rnn.addRecurrentConnection(FullConnection(rnn['hidden1'], rnn['hidden1']))
rnn.addRecurrentConnection(FullConnection(rnn['hidden2'], rnn['hidden2']))
rnn.sortModules()
trainer = RPropMinusTrainer(rnn,
dataset=trainData,
batchlearning=True,
verbose=True,
weightdecay=0.005)
errTime = []
errTrain = []
errValidation = []
epochNo = 0
while True:
for i in range(config['epochs_per_update']):
trainer.train()
epochNo += config['epochs_per_update']
NetworkWriter.writeToFile(rnn, '%s/Epoch_%d.xml' % (logDir, epochNo))
NetworkWriter.writeToFile(rnn, '%s/Latest.xml' % logDir)
tOut = ModuleValidator.calculateModuleOutput(rnn, trainData)
vOut = ModuleValidator.calculateModuleOutput(rnn, validationData)
tScaler = config.getDataScale([config['output_scalar_label']])[0][1]
tAvgErr = NP.sqrt(NP.mean((trainData['target'] - tOut)**2)) * tScaler
vAvgErr = NP.sqrt(NP.mean(
(validationData['target'] - vOut)**2)) * tScaler
tMaxErr = NP.max(NP.abs(trainData['target'] - tOut)) * tScaler
vMaxErr = NP.max(NP.abs(validationData['target'] - vOut)) * tScaler
errTrain.append(tAvgErr)
errValidation.append(vAvgErr)
errTime.append(epochNo)
print "Training error: avg %5.3f max %5.3f" % (tAvgErr,
tMaxErr)
print "Validation error: avg %5.3f max %5.3f" % (vAvgErr,
vMaxErr)
print "------------------------------------------------------------------------------"
if (config['visualize_on_training'] == 'yes'):
PL.figure(1)
PL.ioff()
visulizeDataSet(rnn, trainData, 0,
config['visualized_columns']['input'],
config['visualized_columns']['output'])
PL.ion()
PL.draw()
PL.figure(2)
PL.ioff()
visulizeDataSet(rnn, validationData, 0,
config['visualized_columns']['input'],
config['visualized_columns']['output'])
PL.ion()
PL.draw()
p = PL.figure(3)
PL.ioff()
p.clear()
PL.plot(errTime, errTrain, label='Train')
PL.plot(errTime, errValidation, label='Validation')
PL.legend()
PL.ion()
PL.draw()
def rnn():
# load dataframe from csv file
df = pi.load_data_frame('../../data/NABIL.csv')
# column name to match with indicator calculating modules
# TODO: resolve issue with column name
df.columns = [
'Transactions', 'Traded_Shares', 'Traded_Amount', 'High', 'Low',
'Close'
]
data = df.Close.values
# TODO: write min_max normalization
# normalization
# cp = dataframe.pop(' Close Price')
# x = cp.values
temp = np.array(data).reshape(len(data), 1)
min_max_scaler = preprocessing.MinMaxScaler()
data = min_max_scaler.fit_transform(temp)
# dataframe[' Close Price'] = x_scaled
# prepate sequential dataset for pyBrain rnn network
ds = SequentialDataSet(1, 1)
for sample, next_sample in zip(data, cycle(data[1:])):
ds.addSample(sample, next_sample)
# build rnn network with LSTM layer
# if saved network is available
if (os.path.isfile('random.xml')):
net = NetworkReader.readFrom('network.xml')
else:
net = buildNetwork(1,
20,
1,
hiddenclass=LSTMLayer,
outputbias=False,
recurrent=True)
# build trainer
trainer = RPropMinusTrainer(net, dataset=ds, verbose=True)
train_errors = [] # save errors for plotting later
EPOCHS_PER_CYCLE = 5
CYCLES = 5
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in range(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
train_errors.append(trainer.testOnData())
epoch = (i + 1) * EPOCHS_PER_CYCLE
print("\r epoch {}/{}".format(epoch, EPOCHS), end="")
sys.stdout.flush()
# save the network
NetworkWriter.writeToFile(net, 'network.xml')
print()
print("final error =", train_errors[-1])
predicted = []
for dat in data:
predicted.append(net.activate(dat)[0])
# data = min_max_scaler.inverse_transform(data)
# predicted = min_max_scaler.inverse_transform(predicted)
predicted_array = min_max_scaler.inverse_transform(
np.array(predicted).reshape(-1, 1))
print(predicted_array[-1])
plt.figure()
legend_actual, = plt.plot(range(0, len(data)),
temp,
label='actual',
linestyle='--',
linewidth=2,
c='blue')
legend_predicted, = plt.plot(range(0, len(data)),
predicted_array,
label='predicted',
linewidth=1.5,
c='red')
plt.legend(handles=[legend_actual, legend_predicted])
plt.savefig('error.png')
plt.show()
for ts in train_data:
ds.newSequence()
# Add obsv and next
for t_1, t_2 in zip(ts, ts[1:]):
ds.addSample(t_1, t_2)
# RNN with 1-5-1 architecture: 1 input, 5 hidden, 1 output layer
rnn = buildNetwork(1,
5,
1,
hiddenclass=LSTMLayer,
outputbias=False,
recurrent=True)
# Initialize trainer
trainer = RPropMinusTrainer(rnn, dataset=ds)
# Predefine iterations: epochs & cycles
EPOCHS_PER_CYCLE = 5
CYCLES = 100
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
# Training loop
for i in xrange(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
error = trainer.testOnData()
epoch = (i + 1) * EPOCHS_PER_CYCLE
print("\r Epoch: {}/{} Error: {}".format(epoch, EPOCHS, error), end="")
stdout.flush()
# Save model
trndata = generateNoisySines(50, 40)
trndata._convertToOneOfMany(bounds=[0., 1.])
tstdata = generateNoisySines(50, 20)
tstdata._convertToOneOfMany(bounds=[0., 1.])
# construct LSTM network - note the missing output bias
rnn = buildNetwork(trndata.indim,
5,
trndata.outdim,
hiddenclass=LSTMLayer,
outclass=SoftmaxLayer,
outputbias=False,
recurrent=True)
# define a training method
trainer = RPropMinusTrainer(rnn, dataset=trndata, verbose=True)
# instead, you may also try
##trainer = BackpropTrainer( rnn, dataset=trndata, verbose=True, momentum=0.9, learningrate=0.00001 )
# carry out the training
for i in xrange(100):
trainer.trainEpochs(2)
trnresult = 100. * (1.0 - testOnSequenceData(rnn, trndata))
tstresult = 100. * (1.0 - testOnSequenceData(rnn, tstdata))
print "train error: %5.2f%%" % trnresult, ", test error: %5.2f%%" % tstresult
# just for reference, plot the first 5 timeseries
plot(trndata['input'][0:250, :], '-o')
hold(True)
plot(trndata['target'][0:250, 0])
show()
dataset_8.addSample(current_sample, next_sample)
for current_sample, next_sample in zip(training_data_9, cycle(training_data_9[1:])):
dataset_9.addSample(current_sample, next_sample)
for current_sample, next_sample in zip(training_data_10, cycle(training_data_10[1:])):
dataset_10.addSample(current_sample, next_sample)
for current_sample, next_sample in zip(testing_data, cycle(testing_data[1:])):
dataset_bis.addSample(current_sample, next_sample)
# Initializing the LSTM RNN: 23 nodes in the hidden layer
network = buildNetwork(1, 23, 1, hiddenclass=LSTMLayer, outputbias=False, recurrent=True)
# Training data
trainer = RPropMinusTrainer(network, dataset=dataset, delta0 = 0.01)
trainer_2 = RPropMinusTrainer(network, dataset=dataset_2, delta0 = 0.01)
trainer_3 = RPropMinusTrainer(network, dataset=dataset_3, delta0 = 0.01)
trainer_4 = RPropMinusTrainer(network, dataset=dataset_4, delta0 = 0.01)
trainer_5 = RPropMinusTrainer(network, dataset=dataset_5, delta0 = 0.01)
trainer_6 = RPropMinusTrainer(network, dataset=dataset_6, delta0 = 0.01)
trainer_8 = RPropMinusTrainer(network, dataset=dataset_8, delta0 = 0.01)
trainer_9 = RPropMinusTrainer(network, dataset=dataset_9, delta0 = 0.01)
trainer_10 = RPropMinusTrainer(network, dataset=dataset_10, delta0 = 0.01)
# Initiazlizing storage for the error curves
train_errors = []
train_errors_2 = []
train_errors_3 = []
train_errors_4 = []
def learn(self,
pathdataset=["dstc4_train"],
Pathdataroot="data",
numberOfHiddenUnit=20,
EPOCHS_PER_CYCLE=10,
CYCLES=40,
weightdecayw=0.01):
print "Start learning LSTM, and make dictionary file"
#Construct dictionary: variable name -> corresponding index of element in i/o vector
print "Star make dictionary: variable name -> corresponding index of element in i/o vector"
self.dictOut = {
} #"TOPIC_SLOT_VALUE" -> corresponding index of element
self.dictIn = {
} #"SPEAKER_{val}"or"TOPIC_{val}","WORD_{word}" "BIO_{BIO}", "CLASS_{slot,value}", ""{defined label}-> corresponding index of element
#-target vector dictionary
index = 0
totalNumSlot = 0
for topic in self.tagsets.keys():
for slot in self.tagsets[topic].keys():
totalNumSlot += 1
for value in self.tagsets[topic][slot]:
self.dictOut[topic + "_" + slot + "_" + value] = index
index += 1
print "totalNumSlot:" + str(totalNumSlot)
print "outputSize:" + str(len(self.dictOut.keys()))
#-input dictionry
dataset = []
for pathdat in pathdataset:
dataset.append(
dataset_walker.dataset_walker(pathdat,
dataroot=Pathdataroot,
labels=False))
#--(sub input vector 1) Class features i.e., Slot and value ratio (Similar to base line)
index = 0
for topic in self.tagsets.keys():
for slot in self.tagsets[topic].keys():
if ("CLASS_" + slot) not in self.dictIn:
self.dictIn["CLASS_" + slot] = index
index += 1
for value in self.tagsets[topic][slot]:
if ("CLASS_" + value) not in self.dictIn:
self.dictIn["CLASS_" + value] = index
index += 1
self.TOTALSIZEOFCLASSFeature = index
f = open(self.FileNameofNumClassFeature, "wb")
pickle.dump(self.TOTALSIZEOFCLASSFeature, f)
f.close()
#--(sub input vector 2) Sentence features
if not self.isUseSentenceRepresentationInsteadofBOW:
index = 0
for elemDataset in dataset:
for call in elemDataset:
for (uttr, _) in call:
#General info1 (CLASS; this feature must be rejistered at first)
if ("SPEAKER_" + uttr["speaker"]) not in self.dictIn:
self.dictIn["SPEAKER_" + uttr["speaker"]] = index
index += 1
if ("TOPIC_" + uttr["segment_info"]["topic"]
) not in self.dictIn:
self.dictIn["TOPIC_" +
uttr["segment_info"]["topic"]] = index
index += 1
#General info2
#-BIO
if ("BIO_" + uttr['segment_info']['target_bio']
) not in self.dictIn:
self.dictIn[
"BIO_" +
uttr['segment_info']['target_bio']] = index
index += 1
#BOW
if LSTMWithBOWTracker.isIgnoreUtterancesNotRelatedToMainTask:
if not (uttr['segment_info']['target_bio'] == "O"):
#-BOW
splitedtrans = self.__getRegurelisedBOW(
uttr["transcript"])
for word in splitedtrans:
if ("WORD_" + word) not in self.dictIn:
self.dictIn["WORD_" + word] = index
index += 1
self.TOTALSIZEOFSENTENCEFeature = index
f = open(self.FileNameofNumSentenceFeature, "wb")
pickle.dump(self.TOTALSIZEOFSENTENCEFeature, f)
f.close()
elif self.isUseSentenceRepresentationInsteadofBOW:
index = 0
for i in range(0, LSTMWithBOWTracker.D2V_VECTORSIZE):
self.dictIn[str(index) + "thElemPV"] = index
index += 1
index = 0
for i in range(0, LSTMWithBOWTracker.D2V_VECTORSIZE):
self.dictIn[str(index) + "thAvrWord"] = index
index += 1
assert self.D2V_VECTORSIZE == LSTMWithBOWTracker.D2V_VECTORSIZE, "D2V_VECTORSIZE is restrected to be same over the class"
else:
assert False, "Unexpected block"
#--(sub input vector 3) Features M1s defined
index = 0
if self.isEnableToUseM1sFeature:
rejisteredFeatures = self.__rejisterM1sInputFeatureLabel(
self.tagsets, dataset)
for rFeature in rejisteredFeatures:
assert rFeature not in self.dictIn, rFeature + " already registered in input vector. Use different label name. "
self.dictIn[rFeature] = index
index += 1
self.TOTALSIZEOFM1DEFINEDFeature = index
f = open(self.FileNameofNumM1Feature, "wb")
pickle.dump(self.TOTALSIZEOFM1DEFINEDFeature, f)
f.close()
print "inputSize:" + str(len(self.dictIn.keys()))
assert self.dictIn[
"CLASS_INFO"] == 0, "Unexpected index CLASS_INFO should has value 0"
assert self.dictIn[
"CLASS_Fort Siloso"] == 334, "Unexpected index CLASS_Fort Siloso should has value 334"
assert self.dictIn[
"CLASS_Yunnan"] == 1344, "Unexpected index CLASS_Yunnan should has value 1611"
#--write
fileObject = open('dictInput.pic', 'w')
pickle.dump(self.dictIn, fileObject)
fileObject.close()
fileObject = open('dictOutput.pic', 'w')
pickle.dump(self.dictOut, fileObject)
fileObject.close()
#Build RNN frame work
print "Start learning Network"
#Capability of network is: (30 hidden units can represents 1048576 relations) wherease (10 hidden units can represents 1024)
#Same to Henderson (http://www.aclweb.org/anthology/W13-4073)?
net = buildNetwork(len(self.dictIn.keys()),
numberOfHiddenUnit,
len(self.dictOut.keys()),
hiddenclass=LSTMLayer,
outclass=SigmoidLayer,
outputbias=False,
recurrent=True)
#Train network
#-convert training data into sequence of vector
convDataset = [] #[call][uttr][input,targetvec]
iuttr = 0
convCall = []
for elemDataset in dataset:
for call in elemDataset:
for (uttr, label) in call:
if self.isIgnoreUtterancesNotRelatedToMainTask:
if uttr['segment_info']['target_bio'] == "O":
continue
#-input
convInput = self._translateUtteranceIntoInputVector(
uttr, call)
#-output
convOutput = [0.0] * len(
self.dictOut.keys()) #Occured:+1, Not occured:0
if "frame_label" in label:
for slot in label["frame_label"].keys():
for value in label["frame_label"][slot]:
convOutput[self.dictOut[
uttr["segment_info"]["topic"] + "_" +
slot + "_" + value]] = 1
#-post proccess
if self.isSeparateDialogIntoSubDialog:
if uttr['segment_info']['target_bio'] == "B":
if len(convCall) > 0:
convDataset.append(convCall)
convCall = []
convCall.append([convInput, convOutput])
#print "Converted utterance" + str(iuttr)
iuttr += 1
if not self.isSeparateDialogIntoSubDialog:
if len(convCall) > 0:
convDataset.append(convCall)
convCall = []
#Online learning
trainer = RPropMinusTrainer(net, weightdecay=weightdecayw)
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
#Shuffle order
ds = SequentialDataSet(len(self.dictIn.keys()),
len(self.dictOut.keys()))
datInd = range(0, len(convDataset))
random.shuffle(
datInd
) #Backpropergation already implemeted data shuffling, however though RpropMinus don't.
for ind in datInd:
ds.newSequence()
for convuttr in convDataset[ind]:
ds.addSample(convuttr[0], convuttr[1])
#Evaluation and Train
epoch = (i + 1) * EPOCHS_PER_CYCLE
print "\r epoch {}/{} Error={}".format(
epoch, EPOCHS, trainer.testOnData(dataset=ds))
stdout.flush()
trainer.trainOnDataset(dataset=ds, epochs=EPOCHS_PER_CYCLE)
NetworkWriter.writeToFile(
trainer.module, "LSTM_" + "Epoche" + str(i + 1) + ".rnnw")
NetworkWriter.writeToFile(trainer.module, "LSTM.rnnw")
from pybrain.tools.shortcuts import buildNetwork
from pybrain.structure.modules import LSTMLayer
from pybrain.structure.modules import SigmoidLayer
net = buildNetwork(inputSize,
inputSize,
outputSize,
hiddenclass=LSTMLayer,
outclass=SigmoidLayer,
outputbias=False,
recurrent=True)
from pybrain.supervised import RPropMinusTrainer
from sys import stdout
trainer = RPropMinusTrainer(net, dataset=ds)
train_errors = [] # save errors for plotting later
EPOCHS_PER_CYCLE = 5
CYCLES = int(sys.argv[2])
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
import matplotlib.pyplot as plt
plt.xlabel('Training Epoch')
plt.ylabel('Shooting Error')
plt.ion()
plt.show()
for i in range(CYCLES):
trainer.trainEpochs(EPOCHS_PER_CYCLE)
train_errors.append(trainer.testOnData())
epoch = (i + 1) * EPOCHS_PER_CYCLE
#code for normalize data in ds
i = np.array([d[0] for d in ds])
inorm = np.array([d[0] for d in ds])
i /= np.max(np.abs(i), axis=0)
o = np.array([d[1] for d in ds])
onorm = np.array([d[1] for d in ds])
o /= np.max(np.abs(o), axis=0)
print routes_frommeasur_ids["11000602"] / np.max(np.abs(inorm), axis=0)
#creating new object for normalized data
nds = SupervisedDataSet(1, 1)
for ix in range(len(ds)):
nds.addSample(i[ix], o[ix])
#print routes_frommeasur
#creating net
net = buildNetwork(nds.indim,
3,
nds.outdim,
bias=True,
hiddenclass=TanhLayer)
#training net
trainer = RPropMinusTrainer(net, verbose=True)
trainer.trainOnDataset(nds, 10)
#trainer.testOnData(verbose=True)
p = net.activate(routes_frommeasur_ids["11000602"] /
np.max(np.abs(inorm), axis=0))
print(p[0] * np.max(np.abs(onorm), axis=0)[0])
# Buils a simple LSTM network with 1 input node, 1 output node and 5 LSTM cells
net = buildNetwork(1,
12,
1,
hiddenclass=LSTMLayer,
peepholes=False,
outputbias=False,
recurrent=True)
# net = buildNetwork(1, 1, 1, hiddenclass=LSTMLayer, peepholes = True, outputbias=False, recurrent=True)
# rnn = buildNetwork( trndata.indim, 5, trndata.outdim, hiddenclass=LSTMLayer, outclass=SoftmaxLayer, outputbias=False, recurrent=True)
from pybrain.supervised import RPropMinusTrainer
from sys import stdout
trainer = RPropMinusTrainer(net, dataset=ds, verbose=True)
#trainer.trainUntilConvergence()
train_errors = [] # save errors for plotting later
EPOCHS_PER_CYCLE = 100 # increasing the epochs to 20, decreases accuracy drastically, decreasing epochs is desiredepoch # 5 err = 0.04
CYCLES = 10 # vary the epochs adn the cycles and the LSTM cells to get more accurate results.
EPOCHS = EPOCHS_PER_CYCLE * CYCLES
for i in xrange(CYCLES):
trainer.trainEpochs(
EPOCHS_PER_CYCLE
) # train on the given data set for given number of epochs
train_errors.append(trainer.testOnData())
epoch = (i + 1) * EPOCHS_PER_CYCLE
print("\r epoch {}/{}".format(epoch, EPOCHS), end="")
stdout.flush()
# trainer.setData(ds)
# import random
# random.shuffle(sequences)
# concat_sequences = []
# for sequence in sequences:
# concat_sequences += sequence
# concat_sequences.append(random.randrange(100, 1000000))
# # concat_sequences = sum(sequences, [])
# for j in xrange(len(concat_sequences) - 1):
# ds.addSample(num2vec(concat_sequences[j], nDim), num2vec(concat_sequences[j+1], nDim))
# trainer.train()
net = initializeLSTMnet(nDim, nLSTMcells=50)
net.reset()
ds = SequentialDataSet(nDim, nDim)
trainer = RPropMinusTrainer(net)
trainer.setData(ds)
for _ in xrange(1000):
# Batch training mode
# print "generate a dataset of sequences"
import random
random.shuffle(sequences)
concat_sequences = []
for sequence in sequences:
concat_sequences += sequence
concat_sequences.append(random.randrange(100, 1000000))
for j in xrange(len(concat_sequences) - 1):
ds.addSample(num2vec(concat_sequences[j], nDim),
num2vec(concat_sequences[j + 1], nDim))
trainer.trainEpochs(rptNum)
