util.py

#-*- coding:utf-8 -*-
'''
Created on 2016��5��16��

@author: Administrator
'''
import numpy as np
from bean import nn,autoencoder
# 激活函数
def sigmod(x):
    return 1.0 / (1.0 + np.exp(-x))

#前馈函数
def nnff(nn,x,y):
    layers = nn.layers
    numbers = x.shape[0]
    # 赋予初值
    nn.values[0] = x
    for i in range(1,layers):
        nn.values[i] = sigmod(np.dot(nn.values[i-1],nn.W[i-1])+nn.B[i-1])
    # 最后一层与实际的误差
    nn.error = y - nn.values[layers-1]
    nn.loss = 1.0/2.0*(nn.error**2).sum()/numbers
    return nn     
#BP函数
def nnbp(nn):
    layers = nn.layers;
    #初始化delta
    
    deltas = list();
    for i in range(layers):
        deltas.append(0)
    
    #最后一层的delta为
    deltas[layers-1] = -nn.error*nn.values[layers-1]*(1-nn.values[layers-1])
    #其他层的delta为
    for j in range(1,layers-1)[::-1]:#倒过来
        deltas[j] = np.dot(deltas[j+1],nn.W[j].T)*nn.values[j]*(1-nn.values[j])
    #更新W值
    for k in range(layers-1):
        nn.W[k] -= nn.u*np.dot(nn.values[k].T,deltas[k+1])/(deltas[k+1].shape[0])
        nn.B[k] -= nn.u*deltas[k+1]/(deltas[k+1].shape[0])
    return nn
#对神经网络进行训练
def nntrain(nn,x,y,iterations):
    for i in range(iterations):
        nnff(nn,x,y)
        nnbp(nn)
    return nn
#建立autoencoder框架
def aebuilder(nodes):
    layers = len(nodes)
    ae = autoencoder()
    for i in range(layers-1):
        ae.add_one(nn([nodes[i],nodes[i+1],nodes[i]]))
    
    return ae
        
#训练autoencoder
def aetrain(ae,x,interations):
    elayers = len(ae.encoders)
    for i in range(elayers):
        #单层训练
        ae.encoders[i] = nntrain(ae.encoders[i],x,x,interations)
        #单层训练后，获取该层中间层的值，作为下一层的训练
        nntemp = nnff(ae.encoders[i],x,x)
        x = nntemp.values[1]   
    return ae     
        
#def predict(nn,x):