def predict(theta, X, ninput, nhidden, noutput):
theta1 = np.reshape(theta[0:nhidden*(ninput+1)], [nhidden, ninput + 1])
theta2 = np.reshape(theta[nhidden*(ninput+1):], [noutput, nhidden + 1])
h1 = sigmoid(np.dot(np_extend(X, 1), theta1.T))
h2 = sigmoid(np.dot(np_extend(h1, 1), theta2.T))
return np.argmax(h2, axis = 1)
def bp(theta, ninput, nhidden, noutput, Lambda, X, y):
'''反向传播, 求得theta的梯度, 这里有很多计算是和fp重复的, 原因在于迭代函数
fmin_cg的参数格式要求, 重复的程度很高, 很影响效率
'''
theta1 = np.reshape(theta[0:nhidden*(ninput+1)], [nhidden, ninput + 1])
theta2 = np.reshape(theta[nhidden*(ninput+1):], [noutput, nhidden + 1])
m = X.shape[0]
a1 = np_extend(X, 1)
z2 = np.dot(a1, theta1.T)
a2 = np_extend(sigmoid(z2), 1)
z3 = np.dot(a2, theta2.T)
a3 = sigmoid(z3)
yTmp = np.eye(noutput)
yy = yTmp[y][:]
delta3 = a3 - yy
delta2 = np.dot(delta3, theta2[:, 1:]) * a2[:, 1:] * (1-a2[:, 1:])
theta1_g = np_extend(Lambda / m * theta1[:, 1:])
theta2_g = np_extend(Lambda / m * theta2[:, 1:])
theta1_g += 1.0 / m * np.dot(delta2.T, a1)
theta2_g += 1.0 / m * np.dot(delta3.T, a2)
grad = np.empty(theta.shape)
grad[0:nhidden*(ninput+1)] = np.reshape(theta1_g, nhidden * (ninput + 1))
grad[nhidden*(ninput+1):] = np.reshape(theta2_g, noutput * (nhidden + 1))
return grad
def predict(theta, X, ninput, nhidden, noutput):
theta1 = np.reshape(theta[0:nhidden * (ninput + 1)], [nhidden, ninput + 1])
theta2 = np.reshape(theta[nhidden * (ninput + 1):], [noutput, nhidden + 1])
h1 = sigmoid(np.dot(np_extend(X, 1), theta1.T))
h2 = sigmoid(np.dot(np_extend(h1, 1), theta2.T))
return np.argmax(h2, axis=1)
def bp(theta, ninput, nhidden, noutput, Lambda, X, y):
'''反向传播, 求得theta的梯度, 这里有很多计算是和fp重复的, 原因在于迭代函数
fmin_cg的参数格式要求, 重复的程度很高, 很影响效率
'''
theta1 = np.reshape(theta[0:nhidden * (ninput + 1)], [nhidden, ninput + 1])
theta2 = np.reshape(theta[nhidden * (ninput + 1):], [noutput, nhidden + 1])
m = X.shape[0]
a1 = np_extend(X, 1)
z2 = np.dot(a1, theta1.T)
a2 = np_extend(sigmoid(z2), 1)
z3 = np.dot(a2, theta2.T)
a3 = sigmoid(z3)
yTmp = np.eye(noutput)
yy = yTmp[y][:]
delta3 = a3 - yy
delta2 = np.dot(delta3, theta2[:, 1:]) * a2[:, 1:] * (1 - a2[:, 1:])
theta1_g = np_extend(Lambda / m * theta1[:, 1:])
theta2_g = np_extend(Lambda / m * theta2[:, 1:])
theta1_g += 1.0 / m * np.dot(delta2.T, a1)
theta2_g += 1.0 / m * np.dot(delta3.T, a2)
grad = np.empty(theta.shape)
grad[0:nhidden * (ninput + 1)] = np.reshape(theta1_g,
nhidden * (ninput + 1))
grad[nhidden * (ninput + 1):] = np.reshape(theta2_g,
noutput * (nhidden + 1))
return grad
def fp(theta, ninput, nhidden, noutput, Lambda, X, y):
'''正向传播求得损失函数
'''
theta1 = np.reshape(theta[0:nhidden*(ninput+1)], [nhidden, ninput + 1])
theta2 = np.reshape(theta[nhidden*(ninput+1):], [noutput, nhidden + 1])
m = X.shape[0]
a1 = np_extend(X, 1)
z2 = np.dot(a1, theta1.T)
a2 = np_extend(sigmoid(z2), 1)
z3 = np.dot(a2, theta2.T)
## sigmoid
# a3 = sigmoid(z3)
# yTmp = np.eye(noutput)
# yy = yTmp[y][:]
# J = np.sum(np.sum(-yy*np.log(a3)-(1-yy)*np.log(1-a3))) / m;
# J += 0.5*Lambda/m * (np.sum(np.sum(theta1[:,1:]*theta1[:,1:])) + \
# np.sum(np.sum(theta2[:,1:]*theta2[:,1:])));
# softmax
a3 = softmax(z3)
J = - np.sum(np.log(a3[range(m), y]));
J += 0.5*Lambda/m* (np.sum(np.sum(theta1[:,1:]*theta1[:,1:])) + \
np.sum(np.sum(theta2[:,1:]*theta2[:,1:])));
return J
def fp(theta, ninput, nhidden, noutput, Lambda, X, y):
'''正向传播求得损失函数
'''
theta1 = np.reshape(theta[0:nhidden * (ninput + 1)], [nhidden, ninput + 1])
theta2 = np.reshape(theta[nhidden * (ninput + 1):], [noutput, nhidden + 1])
m = X.shape[0]
a1 = np_extend(X, 1)
z2 = np.dot(a1, theta1.T)
a2 = np_extend(sigmoid(z2), 1)
z3 = np.dot(a2, theta2.T)
## sigmoid
# a3 = sigmoid(z3)
# yTmp = np.eye(noutput)
# yy = yTmp[y][:]
# J = np.sum(np.sum(-yy*np.log(a3)-(1-yy)*np.log(1-a3))) / m;
# J += 0.5*Lambda/m * (np.sum(np.sum(theta1[:,1:]*theta1[:,1:])) + \
# np.sum(np.sum(theta2[:,1:]*theta2[:,1:])));
# softmax
a3 = softmax(z3)
J = -np.sum(np.log(a3[range(m), y]))
J += 0.5*Lambda/m* (np.sum(np.sum(theta1[:,1:]*theta1[:,1:])) + \
np.sum(np.sum(theta2[:,1:]*theta2[:,1:])))
return J