相关算法
python代码参考http://blog.csdn.net/zc02051126/article/details/9668439#（作少量修改与注释）
#coding:utf8
import matplotlib.pylab as plt
import numpy as np
import cPickle class RBM:
def __init__(self,n_visul, n_hidden, max_epoch = 50, batch_size = 110, penalty = 2e-4):
self.n_visible = n_visul
self.n_hidden = n_hidden
self.max_epoch = max_epoch
self.batch_size = batch_size
self.penalty = penalty
self.w = np.random.random((self.n_visible, self.n_hidden)) * 0.1
self.v_bias = np.zeros((1, self.n_visible))
self.h_bias = np.zeros((1, self.n_hidden)) def sigmoid(self, z):
return 1.0 / (1.0 + np.exp( -z )) def forward(self, vis):
return self.sigmoid(np.dot(vis.T, self.w) + self.h_bias) def backward(self, vis):
return self.sigmoid(np.dot(vis, self.w.T) + self.v_bias) def batch(self):
d, N = self.x.shape
num_batchs = int(round(N / self.batch_size)) + 1
groups = np.ravel(np.repeat([range(0, num_batchs)], self.batch_size, axis = 0))
groups=groups[:N]
np.random.shuffle(groups)
batch_data = []
for i in range(0, num_batchs):
index = groups == i
batch_data.append(self.x[:, index])
return batch_data def rbmBB(self, x):
self.x = x
eta = 0.1
momentum = 0.5 #动量项
W = self.w
b = self.h_bias
c = self.v_bias
Winc = np.zeros((self.n_visible, self.n_hidden))
binc = np.zeros(self.n_hidden)
cinc = np.zeros(self.n_visible)
batch_data = self.batch()
num_batch = len(batch_data)
errors = []
for epoch in range(0, self.max_epoch):
err_sum = 0.0
for batch in range(0, num_batch):
num_dims, num_cases = batch_data[batch].shape
data = batch_data[batch]
# 已知可见层，采样出隐藏层
ph = self.forward(data)
ph_states = np.zeros((num_cases, self.n_hidden))
ph_states[ph > np.random.random((num_cases, self.n_hidden))] = 1
# 已知隐藏层，采样出可见层
neg_data = self.backward(ph_states)
neg_data_states = np.zeros((num_cases, num_dims))
neg_data_states[neg_data > np.random.random((num_cases, num_dims))] = 1
neg_data_states = neg_data_states.transpose()
nh = self.forward(neg_data_states)
# CD算法
dW = np.dot(data, ph) - np.dot(neg_data_states, nh)
dc = np.sum(data, axis = 1) - np.sum(neg_data_states, axis = 1)
db = np.sum(ph, axis = 0) - np.sum(nh, axis = 0)
# 刷新参数
Winc = momentum * Winc + eta * (dW / num_cases - self.penalty * W)
binc = momentum * binc + eta * (db / num_cases);
cinc = momentum * cinc + eta * (dc / num_cases);
W = W + Winc
b = b + binc
c = c + cinc
self.w = W
self.h_bais = b
self.v_bias = c
err = np.linalg.norm(data - neg_data.transpose())
err_sum += err
print epoch, err_sum
errors.append(err_sum)
self.errors = errors
self.hiden_value = self.forward(self.x)
h_row, h_col = self.hiden_value.shape
hiden_states = np.zeros((h_row, h_col))
hiden_states[self.hiden_value > np.random.random((h_row, h_col))] = 1
self.rebuild_value = self.backward(hiden_states) def visualize(self, X): #可视化
D, N = X.shape
s = int(np.sqrt(D))
num = int(np.ceil(np.sqrt(N)))
a = np.zeros((num*s + num + 1, num * s + num + 1)) - 1.0
x = 0
y = 0
for i in range(0, N):
z = X[:,i]
z = z.reshape(s,s,order='F')
z = z.transpose()
a[x*s+x:x*s+s+x , y*s+y:y*s+s+y] = z
x = x + 1
if(x >= num):
x = 0
y = y + 1
return a def readData(path):
data = []
for line in open(path, 'r'):
ele = line.split(' ')
tmp = []
for e in ele:
if e != '':
tmp.append(float(e.strip(' ')))
data.append(tmp)
return data if __name__ == '__main__':
f = open('mnist.pkl', 'rb')
training_data, validation_data, test_data = cPickle.load(f)
training_inputs = [np.reshape(x, 784) for x in training_data[0]]
data =training_inputs[:5000]
data = np.array(data)
data = data.transpose()
rbm = Rbm(784, 100,max_epoch = 50)
rbm.rbmBB(data) a = rbm.visualize(data) #(2060L, 2060L)
fig = plt.figure(1)
ax = fig.add_subplot(111)
ax.imshow(a)
plt.title('original data') rebuild_value = rbm.rebuild_value.transpose()
b = rbm.visualize(rebuild_value) #(2060L, 2060L)
fig = plt.figure(2)
ax = fig.add_subplot(111)
ax.imshow(b)
plt.title('rebuild data') hidden_value = rbm.hiden_value.transpose()
c = rbm.visualize(hidden_value) #(782L, 782L)
fig = plt.figure(3)
ax = fig.add_subplot(111)
ax.imshow(c)
plt.title('hidden data') w_value = rbm.w
d = rbm.visualize(w_value) #(291L, 291L)
fig = plt.figure(4)
ax = fig.add_subplot(111)
ax.imshow(d)
plt.title('weight value(w)')
plt.show()