将matlab代码http://code.google.com/p/matrbm/中rbmBB改写成Python代码,如下,参考文献为:
1 A Tutorial on Stochastic Approximation Algorithms for Training Restricted Boltzmann Machines and Deep Belief Nets
2 Inductive Principles for Learning Restricted Boltzmann Machines
3 Training products of experts by minimizing contrastive divergence
4 受限波尔兹曼机简介
- import matplotlib.pylab as plt
- import numpy as np
- import random
- from scipy.linalg import norm
- import PIL.Image
- class Rbm:
- def __init__(self,n_visul, n_hidden, max_epoch = 50, batch_size = 110, penalty = 2e-4, anneal = False, w = None, v_bias = None, h_bias = None):
- self.n_visible = n_visul
- self.n_hidden = n_hidden
- self.max_epoch = max_epoch
- self.batch_size = batch_size
- self.penalty = penalty
- self.anneal = anneal
- if w is None:
- self.w = np.random.random((self.n_visible, self.n_hidden)) * 0.1
- if v_bias is None:
- self.v_bias = np.zeros((1, self.n_visible))
- if h_bias is None:
- self.h_bias = np.zeros((1, self.n_hidden))
- def sigmod(self, z):
- return 1.0 / (1.0 + np.exp( -z ))
- def forward(self, vis):
- #if(len(vis.shape) == 1):
- #vis = np.array([vis])
- #vis = vis.transpose()
- #if(vis.shape[1] != self.w.shape[0]):
- vis = vis.transpose()
- pre_sigmod_input = np.dot(vis, self.w) + self.h_bias
- return self.sigmod(pre_sigmod_input)
- def backward(self, vis):
- #if(len(vis.shape) == 1):
- #vis = np.array([vis])
- #vis = vis.transpose()
- #if(vis.shape[0] != self.w.shape[1]):
- back_sigmod_input = np.dot(vis, self.w.transpose()) + self.v_bias
- return self.sigmod(back_sigmod_input)
- def batch(self):
- eta = 0.1
- momentum = 0.5
- 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[0 : N]
- perm = range(0, N)
- random.shuffle(perm)
- groups = groups[perm]
- 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
- W = self.w
- b = self.h_bias
- c = self.v_bias
- Wavg = W
- bavg = b
- cavg = c
- Winc = np.zeros((self.n_visible, self.n_hidden))
- binc = np.zeros(self.n_hidden)
- cinc = np.zeros(self.n_visible)
- avgstart = self.max_epoch - 5;
- batch_data = self.batch()
- num_batch = len(batch_data)
- oldpenalty= self.penalty
- t = 1
- errors = []
- for epoch in range(0, self.max_epoch):
- err_sum = 0.0
- if(self.anneal):
- penalty = oldpenalty - 0.9 * epoch / self.max_epoch * oldpenalty
- for batch in range(0, num_batch):
- num_dims, num_cases = batch_data[batch].shape
- data = batch_data[batch]
- #forward
- 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
- #backward
- nh_states = ph_states
- neg_data = self.backward(nh_states)
- neg_data_states = np.zeros((num_cases, num_dims))
- neg_data_states[neg_data > np.random.random((num_cases, num_dims))] = 1
- #forward one more time
- neg_data_states = neg_data_states.transpose()
- nh = self.forward(neg_data_states)
- nh_states = np.zeros((num_cases, self.n_hidden))
- nh_states[nh > np.random.random((num_cases, self.n_hidden))] = 1
- #update weight and biases
- 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)
- 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
- if(epoch > avgstart):
- Wavg -= (1.0 / t) * (Wavg - W)
- cavg -= (1.0 / t) * (cavg - c)
- bavg -= (1.0 / t) * (bavg - b)
- t += 1
- else:
- Wavg = W
- bavg = b
- cavg = c
- #accumulate reconstruction error
- err = 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)
- self.w = Wavg
- self.h_bais = b
- self.v_bias = c
- def visualize(self, X):
- D, N = X.shape
- s = int(np.sqrt(D))
- if s == int(np.floor(s)):
- 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+1+x - 1:x*s+s+x , y*s+1+y - 1:y*s+s+y ] = z
- x = x + 1
- if(x >= num):
- x = 0
- y = y + 1
- d = True
- else:
- a = X
- 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__':
- data = readData('data.txt')
- data = np.array(data)
- data = data.transpose()
- rbm = Rbm(784, 100,max_epoch = 50)
- rbm.rbmBB(data)
- a = rbm.visualize(data)
- 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)
- 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)
- 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)
- fig = plt.figure(4)
- ax = fig.add_subplot(111)
- ax.imshow(d)
- plt.title('weight value(w)')
- plt.show()
产生的结果图片如下:
数据解释:
程序中变量data存储了数据,data的维数为784x5000,每一列代表一幅手写数字的图像数据,每一列中包括了784个像素,把这784个像素转化成28X28的矩阵数据,显示出来即可看出对应的数字,以第0列的数据为例,其手写数字为
实现的代码为:
- c = data[:,0]
- d = c.reshape(28,28,order='F')
- d = d.transpose()
- plt.imshow(d)
- plt.show()