def test_UIThread(self):
print "Testing...."
inp_q = Queue.Queue()
out_q = Queue.Queue()
ui = UIThread(inputChannel=inp_q, sigChannel=out_q)
ui.start()
for i in range(6):
time.sleep(1)
msg = {'Speed': float(i), 'Distance': 20*i}
inp_q.put(msg)
ui.exitSelf()
def __init__(self, c, rnn, nepochs=12, alpha=0.1):
self.nepochs = nepochs
self.alpha = alpha
self.c = c
self.rnn = rnn
self.display_q = Queue.Queue()
self.signal_q = Queue.Queue()
self.ui = UIThread(inputChannel=self.display_q,
sigChannel=self.signal_q)
self.ui.start()
info_st = {
"p": "Start/Pause Training",
"y": "Save Params",
"s": "Sentance Testing",
}
self.display_q.put(info_st)
self.scores = {}
def __init__(self, c, rnn, nepochs = 12, alpha = 0.1):
self.nepochs = nepochs
self.alpha = alpha
self.c = c
self.rnn = rnn
self.display_q = Queue.Queue()
self.signal_q = Queue.Queue()
self.ui = UIThread(inputChannel=self.display_q,sigChannel=self.signal_q)
self.ui.start()
info_st = {
"p": "Start/Pause Training",
"y": "Save Params",
"s": "Sentance Testing",
}
self.display_q.put(info_st)
self.scores = {}
def test_UIThread(self):
print "Testing...."
inp_q = Queue.Queue()
out_q = Queue.Queue()
ui = UIThread(inputChannel=inp_q, sigChannel=out_q)
ui.start()
for i in range(6):
time.sleep(1)
msg = {'Speed': float(i), 'Distance': 20 * i}
inp_q.put(msg)
ui.exitSelf()
class Trainer():
def __init__(self, c, rnn, nepochs = 12, alpha = 0.1):
self.nepochs = nepochs
self.alpha = alpha
self.c = c
self.rnn = rnn
self.display_q = Queue.Queue()
self.signal_q = Queue.Queue()
self.ui = UIThread(inputChannel=self.display_q,sigChannel=self.signal_q)
self.ui.start()
info_st = {
"p": "Start/Pause Training",
"y": "Save Params",
"s": "Sentance Testing",
}
self.display_q.put(info_st)
self.scores = {}
def sendMsg(self, msg):
info = {"msg": msg}
self.display_q.put(info)
def mainEventLoop(self):
sig = 0
while (sig != 1 and sig !=2):
sig = self.checkForSignals()
if sig == 2:
self.sendMsg("Training Called")
self.train()
def train(self):
self.sendMsg("Starting Training")
X = self.c.X
Y = self.c.Y
self.lastAvg = 0.0
ecosts = []
for e in xrange(self.nepochs):
lastCost = 0.0
for i in range(len(X)):
xseq = X[i]
yseq = Y[i]
if i < 3:
niter = 4 - i
else:
niter = 1
for k in xrange(niter):
try:
cost = self.rnn.train(list(xseq), list(yseq), self.alpha)
info = {
'Sentance': i,
'Epoch': e,
'Cost': cost,
'Last Average': self.lastAvg,
'Scores': self.scores,
}
self.display_q.put(info)
sig = self.checkForSignals()
if sig == 1:
self.quitTraining()
return
if sig == 2:
sig2 = 0
while sig2 != 2 :
sig2 = self.checkForSignals()
except KeyboardInterrupt:
pass
self.quitTraining()
return
lastCost += cost
avgCost = lastCost / (i+1)
ecosts.append(avgCost)
self.rnn.save("epoch%d"%e)
# self.fullDataTest()
def quitTraining(self):
self.rnn.save("beforequit")
print Fore.RED, "\n\n QUITTING TRAINING...\n\n"
def checkForSignals(self):
try:
msgCh = self.signal_q.get_nowait()
if msgCh == ord('q'):
return 1
if msgCh == ord('p'):
return 2
if msgCh == ord('s'):
strng = self.ui.getTextString()
strng = strng.replace("\n", "")
print strng.split(' ')
score = self.findProbabilityOfSentance(strng)
self.scores[strng] = score
self.display_q.put(self.scores)
if msgCh == ord('t'):
self.calcPerplexicity()
if msgCh == ord('c'):
self.fullDataTest()
if msgCh == ord('y'):
strng = self.ui.getTextString()
strng = strng.replace("\n", "")
self.rnn.save("%s.save"%strng)
except Queue.Empty :
pass
return 0
def fullDataTest(self):
X = self.c.X
Y = self.c.Y
totAcc = 0.0
for i in range(len(X)):
xseq = X[i]
yseq = Y[i]
acc = self.test_sentance(xseq, yseq)
totAcc += acc
avgAcc = totAcc / (i+1)
info = {
'Sentance': i,
'AvgCost': "%.2f %%"%avgAcc,
}
self.display_q.put(info)
self.lastAvg = avgAcc
return totAcc/len(X)
def test_sentance(self,x , y):
ypred = self.rnn.classify(list(x),list(y))
accuracy = float(np.count_nonzero(ypred-y))
accuracy /= float(len(y))
accuracy = 100*(1.0-accuracy)
return accuracy
def generate_sequence(self):
seq = [self.c.wordToNum["<s>"]]
maxLen = 20
for i in range(maxLen):
dummyY = range(len(seq))
ypred = self.rnn.classify(seq, dummyY)
lastWordIdx = list(ypred)[-1]
seq.append(lastWordIdx)
print Fore.GREEN, self.c.idx_to_sentance(seq)
def calcPerplexicity(self, c=None):
if c == None:
X = self.c.X
Y = self.c.Y
else :
X,Y = self.c.get_XY_For_another_corpus(c)
totAcc = 0.0
totWords = 0
totPer = 0.0
minPer = 999999
maxPer = 0
for i in range(len(X)):
xseq = X[i]
yseq = Y[i]
# curP = self.Perplexicity_xy(xseq, yseq)
cost = self.rnn.cost(list(xseq),list(yseq))
curP = np.exp(cost)
totPer += curP
if maxPer < curP:
maxPer = curP
if minPer > curP and curP > 10:
minPer = curP
totAcc += cost * len(xseq)
totWords += len(xseq)
avgAcc = np.exp(totAcc /totWords )
totSumAvg = totPer / (i+1)
info = {
'Sentance': i,
'PERPLEXITY': "%.2f"%avgAcc,
'Current': "%.2f"%curP,
'Min': minPer,
'Max': maxPer,
'SumAvgPer' : totSumAvg,
}
self.display_q.put(info)
return avgAcc
def Perplexicity_xy(self,x,y):
cost = self.rnn.cost(list(x),list(y))
return np.exp2(cost)
def findProbabilityOfSentance(self, sentance = "The president said he will ask congress"):
words = sentance.split(' ')
# words.append("</s>")
# words.insert(0,"<s>")
xseq = self.c.seq_to_indices(words)
y = xseq[1:] #= range(len(xseq))
x = xseq[:-1]
return self.Perplexicity_xy(x,y)
class Trainer():
def __init__(self, c, rnn, nepochs=12, alpha=0.1):
self.nepochs = nepochs
self.alpha = alpha
self.c = c
self.rnn = rnn
self.display_q = Queue.Queue()
self.signal_q = Queue.Queue()
self.ui = UIThread(inputChannel=self.display_q,
sigChannel=self.signal_q)
self.ui.start()
info_st = {
"p": "Start/Pause Training",
"y": "Save Params",
"s": "Sentance Testing",
}
self.display_q.put(info_st)
self.scores = {}
def sendMsg(self, msg):
info = {"msg": msg}
self.display_q.put(info)
def mainEventLoop(self):
sig = 0
while (sig != 1 and sig != 2):
sig = self.checkForSignals()
if sig == 2:
self.sendMsg("Training Called")
self.train()
def train(self):
self.sendMsg("Starting Training")
X = self.c.X
Y = self.c.Y
self.lastAvg = 0.0
ecosts = []
for e in xrange(self.nepochs):
lastCost = 0.0
for i in range(len(X)):
xseq = X[i]
yseq = Y[i]
if i < 3:
niter = 4 - i
else:
niter = 1
for k in xrange(niter):
try:
cost = self.rnn.train(list(xseq), list(yseq),
self.alpha)
info = {
'Sentance': i,
'Epoch': e,
'Cost': cost,
'Last Average': self.lastAvg,
'Scores': self.scores,
}
self.display_q.put(info)
sig = self.checkForSignals()
if sig == 1:
self.quitTraining()
return
if sig == 2:
sig2 = 0
while sig2 != 2:
sig2 = self.checkForSignals()
except KeyboardInterrupt:
pass
self.quitTraining()
return
lastCost += cost
avgCost = lastCost / (i + 1)
ecosts.append(avgCost)
self.rnn.save("epoch%d" % e)
# self.fullDataTest()
def quitTraining(self):
self.rnn.save("beforequit")
print Fore.RED, "\n\n QUITTING TRAINING...\n\n"
def checkForSignals(self):
try:
msgCh = self.signal_q.get_nowait()
if msgCh == ord('q'):
return 1
if msgCh == ord('p'):
return 2
if msgCh == ord('s'):
strng = self.ui.getTextString()
strng = strng.replace("\n", "")
print strng.split(' ')
score = self.findProbabilityOfSentance(strng)
self.scores[strng] = score
self.display_q.put(self.scores)
if msgCh == ord('t'):
self.calcPerplexicity()
if msgCh == ord('c'):
self.fullDataTest()
if msgCh == ord('y'):
strng = self.ui.getTextString()
strng = strng.replace("\n", "")
self.rnn.save("%s.save" % strng)
except Queue.Empty:
pass
return 0
def fullDataTest(self):
X = self.c.X
Y = self.c.Y
totAcc = 0.0
for i in range(len(X)):
xseq = X[i]
yseq = Y[i]
acc = self.test_sentance(xseq, yseq)
totAcc += acc
avgAcc = totAcc / (i + 1)
info = {
'Sentance': i,
'AvgCost': "%.2f %%" % avgAcc,
}
self.display_q.put(info)
self.lastAvg = avgAcc
return totAcc / len(X)
def test_sentance(self, x, y):
ypred = self.rnn.classify(list(x), list(y))
accuracy = float(np.count_nonzero(ypred - y))
accuracy /= float(len(y))
accuracy = 100 * (1.0 - accuracy)
return accuracy
def generate_sequence(self):
seq = [self.c.wordToNum["<s>"]]
maxLen = 20
for i in range(maxLen):
dummyY = range(len(seq))
ypred = self.rnn.classify(seq, dummyY)
lastWordIdx = list(ypred)[-1]
seq.append(lastWordIdx)
print Fore.GREEN, self.c.idx_to_sentance(seq)
def calcPerplexicity(self, c=None):
if c == None:
X = self.c.X
Y = self.c.Y
else:
X, Y = self.c.get_XY_For_another_corpus(c)
totAcc = 0.0
totWords = 0
totPer = 0.0
minPer = 999999
maxPer = 0
for i in range(len(X)):
xseq = X[i]
yseq = Y[i]
# curP = self.Perplexicity_xy(xseq, yseq)
cost = self.rnn.cost(list(xseq), list(yseq))
curP = np.exp(cost)
totPer += curP
if maxPer < curP:
maxPer = curP
if minPer > curP and curP > 10:
minPer = curP
totAcc += cost * len(xseq)
totWords += len(xseq)
avgAcc = np.exp(totAcc / totWords)
totSumAvg = totPer / (i + 1)
info = {
'Sentance': i,
'PERPLEXITY': "%.2f" % avgAcc,
'Current': "%.2f" % curP,
'Min': minPer,
'Max': maxPer,
'SumAvgPer': totSumAvg,
}
self.display_q.put(info)
return avgAcc
def Perplexicity_xy(self, x, y):
cost = self.rnn.cost(list(x), list(y))
return np.exp2(cost)
def findProbabilityOfSentance(
self, sentance="The president said he will ask congress"):
words = sentance.split(' ')
# words.append("</s>")
# words.insert(0,"<s>")
xseq = self.c.seq_to_indices(words)
y = xseq[1:] #= range(len(xseq))
x = xseq[:-1]
return self.Perplexicity_xy(x, y)
