受限制玻尔兹曼机(RBM)用于电影推荐小例

 1 #coding=‘utf-8‘
 2 """
 3 desc: RBM model
 4 author:luchi
 5 date:9/3/17
 6 """
 7 import numpy as np
 8 class RBM_Model(object):
 9 
10     def __init__(self,visible_size,hidden_size,lr):
11         self.visible_size = visible_size
12         self.hidden_size = hidden_size
13         self.lr = lr
14         np.random.seed(10)
15         self.b_v = np.random.uniform(-1,1,size=[self.visible_size])*0
16         self.W = np.random.uniform(-1,1,size=[self.visible_size,self.hidden_size])
17         self.b_h = np.random.uniform(-1,1,size=[self.hidden_size])*0
18 
19     def sampling(self,data):
20 
21         """
22         sampling h_0 using v_0
23         """
24         
25         h_0 = self.logist_fun(np.dot(data,self.W)+self.b_h)
26         #print h_0
27         h_shape = np.shape(h_0)
28         #h_0_state = h_0>(np.random.rand(h_shape[0],h_shape[1]))
29         h_0_state = h_0>(np.ones_like(h_0)*0.5)
30        
31        
32         """
33         building contrastive sampling
34         """
35         v_1 = self.logist_fun(np.dot(h_0_state,np.transpose(self.W))+self.b_v)
36         v_shape = np.shape(v_1)
37         #v_1_state = v_1>(np.random.rand(v_shape[0],v_shape[1]))
38         v_1_state = v_1>(np.ones_like(v_1)*0.5)
39         h_1 = self.logist_fun(np.dot(v_1,self.W)+self.b_h)
40 
41         return h_0,v_1,h_1,v_1_state
42 
43 
44     def train(self,data,iter_time):
45 
46         h_0,v_1,h_1,v_1_state = self.sampling(data)
47         if iter_time%100==0:
48             error = np.sum(np.mean((data-v_1) ** 2,axis=0))
49             print("the %i iter_time error is %s" % (iter_time, error))
50             
51         """
52         updating weight
53         """
54         updating_matrix = []
55         size = len(data)
56 
57         for i in range(size):
58             w_v0= np.reshape(data[i],[self.visible_size,1])
59             w_h0 = np.reshape(h_0[i],[1,self.hidden_size])
60             w_u0 = np.dot(w_v0,w_h0)
61 
62             w_v1 = np.reshape(v_1[i],[self.visible_size,1])
63             w_h1 =  np.reshape(h_1[i],[1,self.hidden_size])
64             w_u1 = np.dot(w_v1,w_h1)
65 
66             updating_matrix.append(w_u0-w_u1)
67         updating_matrix =  np.mean(np.array(updating_matrix),axis=0)
68         self.W =  self.W + self.lr*updating_matrix
69         self.b_v = self.b_v + self.lr*np.mean((data-v_1),axis=0)
70         self.b_h = self.b_h + self.lr*np.mean((h_0-h_1),axis=0)
71 
72     def logist_fun(self,narray):
73         narray =  np.clip(narray,-100,100)
74         return 1.0/(1+np.exp(-1*narray))
75 
76     def softmax(self,narray):
77         narray =  np.clip(narray,-100,100)
78         num_a = np.exp(narray)
79         num_b = np.sum(num_a,axis=1)
80         return num_a*1.0/num_b[:,None]
81 
82     def recomendation(self,test_data,topK):
83        
84         h_0,v_1,h_1 ,_= self.sampling(test_data)
85         sorted_index = np.argsort(-1*v_1,axis=1)
86         return sorted_index[:,:topK]