def find_a_network():
# TODO notes from thesis update
#
# Look at and implement the quasi-random initialization stuff
# from pga's paper.
#
# Try training networks on other chaotic functions
# In particular try mu * ( x * (1 - x) )^(1/2)
# pga -- it looks like the function that my network learned
#
# Try finding chaotic autonomous RNNs using my GA.
#
K = 80
pat = generate_pattern(200)
network1 = NN(2, K, 1)
d = shelve.open('shelved.networks.db')
key = str(int(time.time()))
d[key] = network1
d.close()
print "Shelved pre-training under the key", key
network1.train(pat, True, 0.010)
print "Done training."
d = shelve.open('shelved.networks.db')
key = str(int(time.time()))
d[key] = network1
d.close()
print "Shelved post-training under the key", key
all_diagrams(lambda x : network1.update([x,1]),f)
def main():
print 'Building neural network...'
net = NN(4, 52, 26)
print 'Training neural network'
count = 0
traindata = []
for filename in listdir(CAPTCHA_DIR):
if count % 100 == 0: print count
count += 1
chdir(CAPTCHA_DIR)
proc = Popen(['../split', filename, '../out.jpg', '1', '0', '210'], stdout=PIPE)
vectors = proc.stdout.read()
proc.wait()
if proc.returncode != 0:
continue
for line in vectors.split('\n'):
line = [x for x in line.split(' ') if x]
if not line: continue
letter = ord(line[0].lower()) - ord('a')
line = line[1:]
line = [int(x) for x in line]
letters = []
for i in xrange(26):
if i == letter:
letters.append(100)
else:
letters.append(0)
traindata.append((line, letters))
net.train(traindata, iterations=10)
print 'Finished training network... Saving'
network = {
'ni': net.ni,
'nh': net.nh,
'no': net.no,
'ai': net.ai,
'ah': net.ah,
'ao': net.ao,
'wi': net.wi,
'wo': net.wo,
'ci': net.ci,
'co': net.co
}
json.dump(network, file('../network2.json', 'w'))
print 'Finished dumping to network2.json'
def main():
print 'Building neural network...'
net = NN(4, 52, 26)
print 'Training neural network'
count = 0
traindata = []
for filename in listdir(CAPTCHA_DIR):
if count % 100 == 0: print count
count += 1
chdir(CAPTCHA_DIR)
proc = Popen(['../split', filename, '../out.jpg', '1', '0', '210'],
stdout=PIPE)
vectors = proc.stdout.read()
proc.wait()
if proc.returncode != 0:
continue
for line in vectors.split('\n'):
line = [x for x in line.split(' ') if x]
if not line: continue
letter = ord(line[0].lower()) - ord('a')
line = line[1:]
line = [int(x) for x in line]
letters = []
for i in xrange(26):
if i == letter:
letters.append(100)
else:
letters.append(0)
traindata.append((line, letters))
net.train(traindata, iterations=10)
print 'Finished training network... Saving'
network = {
'ni': net.ni,
'nh': net.nh,
'no': net.no,
'ai': net.ai,
'ah': net.ah,
'ao': net.ao,
'wi': net.wi,
'wo': net.wo,
'ci': net.ci,
'co': net.co
}
json.dump(network, file('../network2.json', 'w'))
print 'Finished dumping to network2.json'
def train(self, session, doc):
# modified bpnn to accept dict as sparse input
(labels, vectors) = doc.get_raw(session)
(nattrs, vectors) = self.renumber_train(session, vectors)
labelSet = set(labels)
lls = len(labelSet)
if lls == 2:
patts = [(vectors[x], [labels[x]]) for x in xrange(len(labels))]
noutput = 1
else:
if lls < 5:
templ = ((0, 0), (1, 0), (0, 1), (1, 1))
elif lls < 9:
templ = ((0, 0, 0), (1, 0, 0), (0, 1, 0), (1, 1, 0), (0, 0, 1),
(1, 0, 1), (0, 1, 1), (1, 1, 1))
else:
# hopefully not more than 16 classes!
templ = ((0, 0, 0, 0), (1, 0, 0, 0), (0, 1, 0, 0),
(1, 1, 0, 0), (0, 0, 1, 0), (1, 0, 1, 0),
(0, 1, 1, 0), (1, 1, 1, 0), (0, 0, 0, 1),
(1, 0, 0, 1), (0, 1, 0, 1), (1, 1, 0, 1),
(0, 0, 1, 1), (1, 0, 1, 1), (0, 1, 1, 1), (1, 1, 1,
1))
rll = range(len(labels))
patts = [(vectors[x], templ[labels[x]]) for x in rll]
noutput = len(templ[0])
# shuffle to ensure not all of class together
r = random.Random()
r.shuffle(patts)
# only way this is at all usable is with small datasets run in psyco
# but is at least fun to play with...
if maxattr * len(labels) > 2000000:
print "Training NN is going to take a LONG time..."
print "Make sure you have psyco enabled..."
n = NN(maxattr, self.hidden, noutput)
self.model = n
n.train(patts, self.iterations, self.learning, self.momentum)
modStr = cPickle.dumps(n)
f = file(mp, 'w')
f.write(modStr)
f.close()
self.predicting = 1
def train(self, session, doc):
# modified bpnn to accept dict as sparse input
(labels, vectors) = doc.get_raw(session)
(nattrs, vectors) = self.renumber_train(session, vectors)
labelSet = set(labels)
lls = len(labelSet)
if lls == 2:
patts = [(vectors[x], [labels[x]]) for x in xrange(len(labels))]
noutput = 1
else:
if lls < 5:
templ = ((0, 0), (1,0), (0, 1), (1, 1))
elif lls < 9:
templ = ((0, 0, 0), (1, 0, 0), (0, 1, 0), (1, 1, 0),
(0, 0, 1), (1, 0, 1), (0, 1, 1), (1, 1, 1))
else:
# hopefully not more than 16 classes!
templ = ((0,0,0,0), (1,0,0,0), (0,1,0,0), (1,1,0,0),
(0,0,1,0), (1,0,1,0), (0,1,1,0), (1,1,1,0),
(0,0,0,1), (1,0,0,1), (0,1,0,1), (1,1,0,1),
(0,0,1,1), (1,0,1,1), (0,1,1,1), (1,1,1,1))
rll = range(len(labels))
patts = [(vectors[x], templ[labels[x]]) for x in rll]
noutput = len(templ[0])
# shuffle to ensure not all of class together
r = random.Random()
r.shuffle(patts)
# only way this is at all usable is with small datasets run in psyco
# but is at least fun to play with...
if maxattr * len(labels) > 2000000:
print "Training NN is going to take a LONG time..."
print "Make sure you have psyco enabled..."
n = NN(maxattr, self.hidden, noutput)
self.model = n
n.train(patts, self.iterations, self.learning, self.momentum)
modStr = cPickle.dumps(n)
f = file(mp, 'w')
f.write(modStr)
f.close()
self.predicting = 1
def load(self, N=.03, M=.007, iterations=5000):
''' Trains the neural network with match statistics for the current heroid '''
patterns = self.data[self.heroid][:300]
print("n = %d" % len(patterns))
NN.train(self, patterns, iterations, N, M)
