(python 3)

 import numpy
from scipy import sparse as S
from matplotlib import pyplot as plt
from scipy.sparse.csr import csr_matrix
import pandas def normalize(x):
V = x.copy()
V -= x.min(axis=1).reshape(x.shape[0],1)
V /= V.max(axis=1).reshape(x.shape[0],1)
return V def sigmoid(x):
#return x*(x > 0)
#return numpy.tanh(x)
return 1.0/(1+numpy.exp(-x)) class RBM():
def __init__(self, n_visible=None, n_hidden=None, W=None, learning_rate = 0.1, weight_decay=1,cd_steps=1,momentum=0.5):
if W == None:
self.W = numpy.random.uniform(-.1,0.1,(n_visible, n_hidden)) / numpy.sqrt(n_visible + n_hidden)
self.W = numpy.insert(self.W, 0, 0, axis = 1)
self.W = numpy.insert(self.W, 0, 0, axis = 0)
else:
self.W=W
self.learning_rate = learning_rate
self.momentum = momentum
self.last_change = 0
self.last_update = 0
self.cd_steps = cd_steps
self.epoch = 0
self.weight_decay = weight_decay
self.Errors = [] def fit(self, Input, max_epochs = 1, batch_size=100):
if isinstance(Input, S.csr_matrix):
bias = S.csr_matrix(numpy.ones((Input.shape[0], 1)))
csr = S.hstack([bias, Input]).tocsr()
else:
csr = numpy.insert(Input, 0, 1, 1)
for epoch in range(max_epochs):
idx = numpy.arange(csr.shape[0])
numpy.random.shuffle(idx)
idx = idx[:batch_size] self.V_state = csr[idx]
self.H_state = self.activate(self.V_state)
pos_associations = self.V_state.T.dot(self.H_state) for i in range(self.cd_steps):
self.V_state = self.sample(self.H_state)
self.H_state = self.activate(self.V_state) neg_associations = self.V_state.T.dot(self.H_state)
self.V_state = self.sample(self.H_state) # Update weights.
w_update = self.learning_rate * ((pos_associations - neg_associations) / batch_size)
total_change = numpy.sum(numpy.abs(w_update))
self.W += self.momentum * self.last_change + w_update
self.W *= self.weight_decay self.last_change = w_update RMSE = numpy.mean((csr[idx] - self.V_state)**2)**0.5
self.Errors.append(RMSE)
self.epoch += 1
print("Epoch %s: RMSE = %s; ||W||: %6.1f; Sum Update: %f" % (self.epoch, RMSE, numpy.sum(numpy.abs(self.W)), total_change))
return self def learning_curve(self):
plt.ion()
#plt.figure()
plt.show()
E = numpy.array(self.Errors)
plt.plot(pandas.rolling_mean(E, 50)[50:]) def activate(self, X):
if X.shape[1] != self.W.shape[0]:
if isinstance(X, S.csr_matrix):
bias = S.csr_matrix(numpy.ones((X.shape[0], 1)))
csr = S.hstack([bias, X]).tocsr()
else:
csr = numpy.insert(X, 0, 1, 1)
else:
csr = X
p = sigmoid(csr.dot(self.W))
p[:,0] = 1.0
return p def sample(self, H, addBias=True):
if H.shape[1] == self.W.shape[0]:
if isinstance(H, S.csr_matrix):
bias = S.csr_matrix(numpy.ones((H.shape[0], 1)))
csr = S.hstack([bias, H]).tocsr()
else:
csr = numpy.insert(H, 0, 1, 1)
else:
csr = H
p = sigmoid(csr.dot(self.W.T))
p[:,0] = 1
return p if __name__=="__main__":
data = numpy.random.uniform(0,1,(100,10))
rbm = RBM(10,15)
rbm.fit(data,1000)
rbm.learning_curve()

Boltzmann机神经网络python实现的更多相关文章

  1. Boltzmann机

    博客园不能上传附件,所以这里贴两张流程图吧.一个是模拟退火算法的流程图(Boltzmann机本实上就是反复退火的过程), 个是Boltzmann调整权值的过程.

  2. 限制Boltzmann机(Restricted Boltzmann Machine)

    起源:Boltzmann神经网络 Boltzmann神经网络的结构是由Hopfield递归神经网络改良过来的,Hopfield中引入了统计物理学的能量函数的概念. 即,cost函数由统计物理学的能量函 ...

  3. 实现一个单隐层神经网络python

    看过首席科学家NG的深度学习公开课很久了,一直没有时间做课后编程题,做完想把思路总结下来,仅仅记录编程主线. 一 引用工具包 import numpy as np import matplotlib. ...

  4. UR机器人通信--上位机通信(python)

    一.通信socket socket()函数 Python 中,我们用 socket()函数来创建套接字,语法格式如下: socket.socket([family[, type[, proto]]]) ...

  5. 运维堡垒机(跳板机)系统 python

    相信各位对堡垒机(跳板机)不陌生,为了保证服务器安全,前面加个堡垒机,所有ssh连接都通过堡垒机来完成,堡垒机也需要有 身份认证,授权,访问控制,审计等功能,笔者用Python基本实现了上述功能. A ...

  6. 机器学习作业(三)多类别分类与神经网络——Python(numpy)实现

    题目太长了!下载地址[传送门] 第1题 简述:识别图片上的数字. import numpy as np import scipy.io as scio import matplotlib.pyplot ...

  7. 径向基(RBF)神经网络python实现

    from numpy import array, append, vstack, transpose, reshape, \ dot, true_divide, mean, exp, sqrt, lo ...

  8. 手写神经网络Python深度学习

    import numpy import scipy.special import matplotlib.pyplot as plt import scipy.misc import glob impo ...

  9. 六.随机神经网络Boltzmann(玻尔兹曼机)

    Hopfield网络具有最优计算功能,然而网络只能严格按照能量函数递减方式演化,很难避免伪状态的出现,且权值容易陷入局部极小值,无法收敛于全局最优解. 如果反馈神经网络的迭代过程不是那么死板,可以在一 ...

随机推荐

  1. Ehcache.xml 配置及属性说明

    1.配置样例 <?xml version="1.0" encoding="UTF-8"?> <ehcache> <diskStor ...

  2. Vue2学习笔记:数据交互vue-resource

    基本语法 必须引入一个库:vue-resource github地址 // 基于全局Vue对象使用http Vue.http.get('/someUrl', [options]).then(succe ...

  3. [翻译] USING GIT IN XCODE [2] 在XCODE中使用GIT[2]

    USING GIT IN XCODE http://www.cimgf.com/2013/12/10/using-git-in-xcode/ USING AN EXISTING REMOTE PROJ ...

  4. url用法

    url中的name用法: 0.定义主rul.py urlpatterns = [ url(r'^sinfors/', include('sinfors.urls', namespace="s ...

  5. docker 自制CentOS 6-lnp镜像

    环境准备 1台centos 6.5镜像虚拟机  febootstrap.docker febootstrap 安装 yum install -y yum-priorities && r ...

  6. 【Alpha】事后诸葛亮

    一. 项目的预期计划 / 项目的现实进展 详见Alpha冲刺博客第一篇 二. 完成项目过程中的体会 详见Alpha冲刺博客第十二篇 三. 团队成员的分工及在Alpha阶段的工作量比例 成员 职务 博客 ...

  7. NSURLProtocol总结:NSURLProtocol 的本质是对特殊的scechme进行特殊的协议定制

    NSURLProtocol 的本质是对特殊的scechme进行特殊的协议定制: 网络(应用层)请求的统一入口是nsurlconnection和nsurlsession; http.htp.mail等协 ...

  8. ansible-playbook快速入门

    一.yaml语法: 1. yaml语法编写 1.1 同层级的字段通过相同缩进表示 1.2 map结构里面key/value用‘:’来分隔 1.3 key/value可以同行写,也可以换行写,换行写必须 ...

  9. 看到了一个pipeline例子,

    pipeline { agent any options { timestamps() } parameters { string(name: 'GIT_BRANCH', defaultValue: ...

  10. etherlime-2-Etherlime Library API-deployer

    Etherlime Library API 库API Deployer部署者 Deployer functionality The main functionality the deployer ex ...