def train(eta, iters):
for it in xrange(iters):
i = random.randint(0, len(text)/seq_len)
j = i * seq_len
X = text[j:(j+seq_len)]
Y = text[(j+1):(j+1+seq_len)]
print "iteration: %s, cost: %s" % (str(it), str(rnn.train(one_hot(X), one_hot(Y), eta, 1.0)))
def train(eta, iters):
for it in xrange(iters):
i = random.randint(0, len(text) / seq_len)
j = i * seq_len
X = text[j:(j + seq_len)]
Y = text[(j + 1):(j + 1 + seq_len)]
print "iteration: %s, cost: %s" % (
str(it), str(rnn.train(one_hot(X), one_hot(Y), eta, 1.0)))
def non_personalized_train(dnodex, eta, iters, ntusers):
for it in xrange(iters):
i = random.randint(0,len(dnodex.plist)-1)
while len(dnodex.plist[i])<=2 or i in dnodex.test_track:
i=random.randint(0,len(dnodex.plist)-1)
X = dnodex.plist[i][:-1]
Y = dnodex.plist[i][1:]
lossf=str(rnn.train(one_hot(X,len(format(dnodex.npoi,'b'))), one_hot(Y,len(format(dnodex.npoi,'b'))), eta, 1.0))
if it%500==0:
print "iteration: %s, cost: %s" % (str(it), lossf)
def infer_stochastic(rnn, k, temperature, start_char=" "):
x = [one_hot(start_char).flatten()]
for i in xrange(k):
probs = rnn.predict_char(x, temperature)
p = np.asarray(probs[0], dtype="float64")
p /= p.sum()
sample = np.random.multinomial(1, p)
sys.stdout.write(one_hot_to_string(sample))
x = [sample]
rnn.reset_state()
def infer_stochastic(rnn, k, temperature, start_char=" "):
x = [one_hot(start_char).flatten()]
for i in xrange(k):
probs = rnn.predict_char(x, temperature)
p = np.asarray(probs[0], dtype="float64")
p /= p.sum()
sample = np.random.multinomial(1, p)
sys.stdout.write(one_hot_to_string(sample))
x = [sample]
rnn.reset_state()
def infer_stochastic(dnodex, rnn):
precision=0
test_case=0.0000001
for test_index in dnodex.test_track:
test=dnodex.plist[test_index]
if test_index%100==0:
sys.stdout.write('\r%d prediction finished, current precision: %4f,%f,%f...' % (test_index,precision/test_case, precision, test_case))
sys.stdout.flush()
if test_index>=500:
break
if len(test)==1:
continue
for index in range(len(test)-1):
x = [one_hot([test[index]],len(format(dnodex.npoi,'b'))).flatten()]
probs = rnn.predict_char(x, 1)
p = np.asarray(probs[0], dtype="float64")
p /= p.sum()
sample = np.random.multinomial(len(p), p)
res=one_hot_to_string(sample)
if res==test[index+1]:
precision+=1.0
test_case+=1.0
rnn.reset_state()
print 'Precision: ', precision/test_case
