def weightSelecter(self, train=False):
classi = self.classifier
classi.weight = (1 / float(classi.actualOutput.shape[0])) * np.ones(
shape=(classi.actualOutput.shape[0], 1))
best = 5.0
for i in range(0, self.maxIteration):
for j in range(0, len(classi.layers)):
classi.layers[j].cofficient = np.random.rand(
classi.layers[j].cofficient.shape[0],
classi.layers[j].cofficient.shape[1])
classi.findEstimates()
if classi.cost < best:
best = classi.cost
self.prediction = classi.predict(self.test)
classifierWeight = np.log((2.5 - classi.cost) / classi.cost)
self.classifierWeight.append(classifierWeight)
costMatrix = (self.classifier.analyseObservation(
self.trainCopy))['cost'].as_matrix()
y = sigmoid(classifierWeight * (costMatrix.reshape(
(costMatrix.shape[0], 1)) > 0.55).astype(int))
if (np.isnan(y)).any(): print "weight comes out to be nan"
newWeight = classi.weight * y
classi.weight = newWeight / sum(newWeight)
print best, classi.input.shape[0]
return self.prediction
def costMatrix(p):
mid = len(p) / 2
p1 = p[0:mid]
p2 = p[mid:2 * mid]
prob = np.transpose(np.matrix(np.vstack((p1, p2))))
a = (sigmoid(np.dot(data, prob)))
b = 1 - np.sum(a, axis=1)
probability = np.matrix(np.hstack((a, b)))
logp = np.log(probability)
cost = -np.sum(np.multiply(target, logp), axis=1)
return cost
def sum1(p):
mid = len(p) / 2
p1 = p[0:mid]
p2 = p[mid:2 * mid]
prob = np.transpose(np.matrix(np.vstack((p1, p2))))
t = np.squeeze(
np.array(
np.ones(shape=(data.shape[0], 1)) -
np.sum((sigmoid(np.dot(data, prob))), axis=1)))
#print t
return t
def nniterate(self, train=False):
classi = self.classifier
classi.weight = (1 / float(classi.actualOutput.shape[0])) * np.ones(
shape=(classi.actualOutput.shape[0], 1))
for i in range(0, self.maxIteration):
for j in range(0, len(classi.layers)):
classi.layers[j].cofficient = np.random.rand(
classi.layers[j].cofficient.shape[0],
classi.layers[j].cofficient.shape[1])
classi.findEstimates()
classifierWeight = np.log((2.5 - classi.cost) / classi.cost)
tes = classi.predict(self.test)
self.prediction = tes * classifierWeight + self.prediction
self.classifierWeight.append(classifierWeight)
costMatrix = (self.classifier.analyseObservation(
self.trainCopy))['cost'].as_matrix()
y = sigmoid(classifierWeight * (costMatrix.reshape(
(costMatrix.shape[0], 1)) > 0.55).astype(int))
if (np.isnan(y)).any(): print "weight comes out to be nan"
newWeight = classi.weight * y
classi.weight = newWeight / sum(newWeight)
self.prediction = self.prediction.div(np.sum(self.prediction, axis=1),
axis=0)
pred = self.prediction
if train:
self.trainCopy.index = self.prediction.index
pred = self.prediction.join(self.trainCopy,
how='left',
rsuffix='_t')
pred['cost'] = np.log(pred['high']) * pred['high_t'] + np.log(
pred['medium']) * pred['medium_t'] + np.log(
pred['low']) * pred['low_t']
print np.sum(pred['cost']
) / self.trainCopy.shape[0], self.trainCopy.shape[0]
return pred
np.array(
np.ones(shape=(data.shape[0], 1)) -
np.sum((sigmoid(np.dot(data, prob))), axis=1)))
#print t
return t
cons = ({
'type': 'ineq',
'fun': lambda p: sum1(p)
}, {
'type':
'ineq',
'fun':
lambda p: np.squeeze(
np.array(sigmoid(np.dot(data, p[0:len(var)].reshape(len(var), 1)))))
}, {
'type':
'ineq',
'fun':
lambda p: np.squeeze(
np.array(
sigmoid(np.dot(data, p[len(var):2 * len(var)].reshape(len(var), 1))
)))
})
# t= np.random.rand(len(var),3)
# r=np.sum(sigmoid(np.dot(data,t)),axis=1,keepdims=True)-np.ones(shape=(data.shape[0],1))
pinitialize = np.random.rand(len(var) * 2)
pinitialize[0] = -6.9
pinitialize[0 + len(var)] = -6.9