def test(self, test_data, numClass): numSample = test_data.shape[0] testingData = test_data[:, 1:] testingLabel = test_data[:, 0:1] prediction = np.zeros((numSample,1)) for i in range(0, numSample): prediction[i,0] = self.predict(testingData[i,:], numClass) predictMat = oneOfK(prediction, numClass) labelMat = oneOfK(testingLabel, numClass) predictionAnalysis(predictMat, labelMat)
def test(self, test_data, numClass, showAnalysis): numSample = test_data.shape[0] testingData = test_data[:, 1:] testingLabel = test_data[:, 0:1] prediction = np.zeros((numSample,1)) for i in range(0, numSample): prediction[i,0] = self.predict(testingData[i,:]) predictMat = oneOfK(prediction, numClass) labelMat = oneOfK(testingLabel, numClass) self.error = np.sum(np.abs(predictMat - labelMat)) / 2 if showAnalysis == 1: predictionAnalysis(predictMat, labelMat)
def test(self, test_data, numClass):
numSample = test_data.shape[0]
self.flow = [test_data[:,1:]]
for i in range(0, len(self.architecture)):
self.flow.append(1 / ( 1 + np.exp(-np.dot(self.flow[i], self.architecture[i][0]) - np.matlib.repmat(self.architecture[i][1], numSample, 1))))
predictMat = getPrediction(self.flow[-1])
labelMat = oneOfK(test_data[:,0],numClass)
#Compute presicion
predictionAnalysis(predictMat, labelMat)
def train(self, training_data, numClass):
numpy.random.seed(2)
numSample = training_data.shape[0]
if 'architecture' not in self.__dict__:
self.architecture = []
for i in self.params['arch']:
uniformRange = 0.2
if not len(self.architecture) == 0:
self.architecture.append([np.random.uniform(0, uniformRange, (self.architecture[-1][0].shape[1],i)) - 0.5 * uniformRange, np.zeros((1,i))])
else:
self.architecture.append([np.random.uniform(0, uniformRange, (training_data.shape[1] - 1,i)) - 0.5 * uniformRange, np.zeros((1,i))])
for p in range(0, 500):
self.flow = [training_data[:,1:]]
self.delta = []
#Forwardpropagation
for i in range(0, len(self.architecture)):
et = np.exp(np.dot(self.flow[i], self.architecture[i][0]) + np.matlib.repmat(self.architecture[i][1], numSample, 1))
self.flow.append(et / ( 1 + et))
delta = self.flow[len(self.architecture)] - oneOfK(training_data[:,0:1], numClass)
df = np.multiply(self.flow[len(self.architecture)], 1 - self.flow[len(self.architecture)])
temp = np.multiply(delta, df)
self.delta = []
for i in range(0, len(self.architecture)):
self.delta.append(0)
self.delta[len(self.architecture) - 1] = temp
#Compute delta of each layer
currentLayer = len(self.architecture) - 2
while currentLayer >= 0:
self.delta[currentLayer] = np.dot(self.delta[currentLayer + 1], self.architecture[currentLayer + 1][0].T)
self.delta[currentLayer] = np.multiply(self.delta[currentLayer], np.multiply(self.flow[currentLayer + 1], 1 - self.flow[currentLayer + 1]))
currentLayer -= 1
#Compute Gradient of each sample
self.dW = []
self.db = []
for j in range(0, len(self.architecture)):
tempDW = []
for k in range(0, numSample):
tempDW.append(np.dot(self.flow[j][k:(k+1),:].T, self.delta[j][k:(k+1),:]))
self.dW.append(tempDW)
self.db.append(self.delta[j])
#Update weights and biases
r = 0.05
rho = 0
if self.momentum == 1:
rho = 0.1
for k in range(0, len(self.architecture)):
tempW = 0
tempB = 0
for kk in range(0, numSample):
tempW = rho * tempW - r * self.dW[k][kk]
tempB = rho * tempB - r * self.db[k][kk:(kk+1),:]
self.architecture[k][0] += tempW
self.architecture[k][1] += tempB
