def gradient(self, x, i):
# 嵌入参数
v, h, w = self.__num_visual, self.__num_hidden, self.__num_hidden * self.__num_visual
self.__weight_v2h = x[0:w].reshape((h, v))
self.__hidden_bias = x[w:(w + h)]
self.__visual_bias = x[(w + h):]
d, s, m = self.points.shape
i = i % m
minibatch = self.points[:, :, i] # 从数据集中取一个minibatch
h_bias = np.tile(self.__hidden_bias, (1, s))
v_bias = np.tile(self.__visual_bias, (1, s))
h_field_0 = sigmoid(self.__weight_v2h.dot(minibatch) + h_bias)
h_state_0 = sample(h_field_0)
v_field_1 = sigmoid(self.__weight_v2h.T.dot(h_state_0) + v_bias)
v_state_1 = v_field_1
h_field_1 = sigmoid(self.__weight_v2h.dot(v_state_1) + h_bias)
gw = (h_field_0.dot(minibatch.T) - h_field_1.dot(v_state_1.T)) / s
gh = (h_field_0 - h_field_1).dot(np.ones((s, 1))) / s
gv = (minibatch - v_state_1).dot(np.ones((s, 1))) / s
weight_cost = 1e-4
cw = weight_cost * self.__weight_v2h
g = np.vstack(((gw - cw).reshape((-1, 1)), gh, gv))
return -g
def compute(self, points, L=None):
"""多层感知器的计算过程"""
if L is None:
L = len(self.b) # 计算层数
D, N = points.shape
y = [] # 初始化输出
for l in range(L):
y.append(sigmoid(self.w[l].dot(points) + np.tile(self.b[l], (1, N))))
points = y[l]
return y
def ffobject(self, x, i):
# 嵌入参数
v, h, w = self.__num_visual, self.__num_hidden, self.__num_hidden * self.__num_visual
self.__weight_v2h = x[0:w].reshape((h, v))
self.__hidden_bias = x[w:(w + h)]
self.__visual_bias = x[(w + h):]
d, s, m = self.points.shape
i = i % m
minibatch = self.points[:, :, i]
h_bias = self.__hidden_bias.dot(np.ones((1, s)))
v_bias = self.__visual_bias.dot(np.ones((1, s)))
# 计算重建误差
h_field_0 = sigmoid(self.__weight_v2h.dot(minibatch) + h_bias)
h_state_0 = sample(h_field_0)
v_field_1 = sigmoid(self.__weight_v2h.T.dot(h_state_0) + v_bias)
y = ((v_field_1 - minibatch) ** 2).sum() / s # 计算在整个minibatch上的平均重建误差
return y
def posterior(self, v_state):
# posterior 计算后验概率
# 在给定显层神经元取值的情况下，计算隐神经元的激活概率
h_field = sigmoid(self.foreward(v_state))
return h_field
def likelihood(self, h_state):
# likelihood 计算似然概率
# 在给定隐层神经元取值的情况下，计算显神经元的激活概率
v_field = sigmoid(self.backward(h_state))
return v_field