-
Notifications
You must be signed in to change notification settings - Fork 0
/
rbm.py
160 lines (107 loc) · 5.48 KB
/
rbm.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Fri Mar 9 21:56:47 2018
@author: baifrank
"""
import torch
##建立RBM的类，声明模型中所有的参数，参数更新公式
class RBM(object):
def __init__(self, num_visible, num_hidden,k=5):
self.dim = (k*num_visible, num_hidden)
self.num_visible = num_visible
self.num_hidden = num_hidden
self.k = k
def params_init(self):
##torch.randn:从N(0,1)中随机抽取
self.weights = torch.randn(self.dim)
self.vbias = torch.randn(1, self.k*self.num_visible)
self.hbias = torch.randn(1, self.num_hidden)
self.D = torch.randn(self.num_visible,self.num_hidden)##conditional RBM的权重
def sample_hidden(self, vis):
wx = torch.mm(vis, self.weights)
activation = wx + self.hbias.expand_as(wx)
p_h_given_v = torch.sigmoid(activation)
##torch.rand:从[0,1]均匀分布中抽取一组随机数
h1_sample = torch.sign(p_h_given_v-torch.rand(p_h_given_v.shape))##如果是nan那么sign之后为0
h1_sample[h1_sample == -1] = 0
return p_h_given_v, h1_sample
##条件RBM抽取隐藏层的函数
def sample_hidden_c(self, vis, r_input):
wx = torch.mm(vis, self.weights)
activation = wx + self.hbias.expand_as(wx) + torch.mm(r_input, self.D)
p_h_given_v = torch.sigmoid(activation)
h1_sample = torch.sign(p_h_given_v-torch.rand(p_h_given_v.shape))
h1_sample[h1_sample == -1] = 0
return p_h_given_v, h1_sample
def sample_visible(self, hid):
wy = torch.mm(hid, self.weights.t())
activation = wy + self.vbias.expand_as(wy)
activation_exp = torch.exp(activation).view(-1,self.k)
##如果exp之后出现0值，则在下面的概率中直接赋一个比较小的值
activation_exp[activation_exp == 0] = 0.01
##将其中inf值变成矩阵中的最大值
max_activation_exp = activation_exp[activation_exp != float('inf')].max()
activation_exp[activation_exp == float('inf')] = max_activation_exp
activation_exp_sum = torch.mm(activation_exp,torch.ones(self.k,self.k).view(-1,self.k))
max_activation_exp_sum = activation_exp_sum[activation_exp_sum != float('inf')].max()
activation_exp_sum[activation_exp_sum == float('inf')] = max_activation_exp_sum
activation_exp_sum[activation_exp_sum == 0] = 0.05
##如果相除之后出现inf值，则在下面的概率中直接赋值为1
p_v_given_h = torch.div(activation_exp,activation_exp_sum).view(-1,self.k*self.num_visible)
p_v_given_h[p_v_given_h == float('inf')] = 1
v1_sample = torch.sign(p_v_given_h-torch.rand(p_v_given_h.shape))
v1_sample[v1_sample == -1] = 0
return p_v_given_h,v1_sample
##没有k，纯粹binary进行抽样
def sample_visible_binary(self, hid):
wy = torch.mm(hid, self.weights.t())
activation = wy + self.vbias.expand_as(wy)
p_v_given_h = torch.sigmoid(activation)
v1_sample = torch.sign(p_v_given_h-torch.rand(p_v_given_h.shape))
v1_sample[v1_sample == -1] = 0
return p_v_given_h,v1_sample
##来回完成一次从v0->h0->v1的采样
def sample_vhv(self,v0):
_, h0 = self.sample_hidden(v0)
pv1, v1 = self.sample_visible(h0)
return v1,pv1
##没有k，纯粹的bianry抽样
def sample_vhv_binary(self,v0):
_, h0 = self.sample_hidden(v0)
pv1, v1 = self.sample_visible_binary(h0)
return v1,pv1
def gradient(self, v0, vk, ph0, phk):
gw = torch.mm(v0.t(), ph0) - torch.mm(vk.t(), phk)
gbv = torch.sum((v0 - vk), 0)
gbh = torch.sum((ph0 - phk), 0)
return gw,gbv,gbh
def train(self, v0, vk, ph0, phk, prev_gw, prev_gbv, prev_gbh, w_lr=0.001,v_lr=0.001,h_lr=0.001,decay=0,momentum=0.9):
##添加学习率衰减和动量系数
gw, gbv, gbh = self.gradient(v0, vk, ph0, phk)
if decay:
gw -= decay * self.weights
gbv -= decay * self.vbias
gbh -= decay * self.hbias
self.weights += momentum * prev_gw + w_lr * gw
self.vbias += momentum * prev_gbv + v_lr * gbv
self.hbias += momentum * prev_gbh + h_lr * gbh
##条件RBM的梯度公式和参数更新公式
def gradient_c(self, v0, vk, ph0, phk, r_input):
gw = torch.mm(v0.t(), ph0) - torch.mm(vk.t(), phk)
gbv = torch.sum((v0 - vk), 0)
gbh = torch.sum((ph0 - phk), 0)
gD = torch.mm(r_input.t(), ph0 - phk )
return gw,gbv,gbh,gD
def train_c(self, v0, vk, ph0, phk, r_input, prev_gw, prev_gbv, prev_gbh, prev_gD, w_lr=0.001,v_lr=0.001,h_lr=0.001,D_lr=0.001,decay=0,momentum=0.9):
##添加学习率衰减和动量系数
gw, gbv, gbh, gD = self.gradient_c(v0, vk, ph0, phk, r_input)
if decay:
gw -= decay * self.weights
gbv -= decay * self.vbias
gbh -= decay * self.hbias
gD -= decay * self.D
self.weights += momentum * prev_gw + w_lr * gw
self.vbias += momentum * prev_gbv + v_lr * gbv
self.hbias += momentum * prev_gbh + h_lr * gbh
self.D += momentum * prev_gD + D_lr * gD