上周末利用python简单实现了一个卷积神经网络,只包含一个卷积层和一个maxpooling层,pooling层后面的多层神经网络采用了softmax形式的输出AI爱发猫。
实验输入仍然采用MNIST图像使用10个featuremap时,卷积和pooling的结果分别如下所示。
部分源码如下:[python] viewplain copy#coding=utf-8'''''Created on 2014年11月30日@author: Wangliaofan'''import numpyimport structimport matplotlib.pyplot as pltimport mathimport randomimport copy#testfrom BasicMultilayerNeuralNetwork import BMNN2def sigmoid(inX):if (-inX)== 0.0:return 999999999.999999999return 1.0/((-inX))def difsigmoid(inX):return sigmoid(inX)*(1.0-sigmoid(inX))def tangenth(inX):return (1.0*(inX)-1.0*(-inX))/(1.0*(inX)+1.0*(-inX))def cnn_conv(in_image, filter_map,B,type_func='sigmoid'):#in_image[num,feature map,row,col]=>in_image[Irow,Icol]#features map[k filter,row,col]#type_func['sigmoid','tangenth']#out_feature[k filter,Irow-row+1,Icol-col+1]shape_image=numpy.shape(in_image)#[row,col]#print "shape_image",shape_imageshape_filter=numpy.shape(filter_map)#[k filter,row,col]if shape_filter[1]>shape_image[0] or shape_filter[2]>shape_image[1]:raise Exceptionshape_out=(shape_filter[0],shape_image[0]-shape_filter[1]+1,shape_image[1]-shape_filter[2]+1)out_feature=numpy.zeros(shape_out)k,m,n=numpy.shape(out_feature)for k_idx in range(0,k):#rotate 180 to calculate convc_filter=numpy.rot90(filter_map[k_idx,:,:], 2)for r_idx in range(0,m):for c_idx in range(0,n):#conv_temp=numpy.zeros((shape_filter[1],shape_filter[2]))(in_image[r_idx:r_idx+shape_filter[1],c_idx:c_idx+shape_filter[2]],c_filter)(conv_temp)if type_func=='sigmoid':out_feature[k_idx,r_idx,c_idx]=sigmoid(sum_temp+B[k_idx])elif type_func=='tangenth':out_feature[k_idx,r_idx,c_idx]=tangenth(sum_temp+B[k_idx])else:raise Exceptionreturn out_featuredef cnn_maxpooling(out_feature,pooling_size=2,type_pooling="max"):k,row,col=numpy.shape(out_feature)max_index_Matirx=numpy.zeros((k,row,col))out_row=int(numpy.floor(row/pooling_size))out_col=int(numpy.floor(col/pooling_size))out_pooling=numpy.zeros((k,out_row,out_col))for k_idx in range(0,k):for r_idx in range(0,out_row):for c_idx in range(0,out_col):temp_matrix=out_feature[k_idx,pooling_size*r_idx:pooling_size*r_idx+pooling_size,pooling_size*c_idx:pooling_size*c_idx+pooling_size]out_pooling[k_idx,r_idx,c_idx](temp_matrix)max_index=numpy.argmax(temp_matrix)#print max_index#print max_index/pooling_size,max_index%pooling_sizemax_index_Matirx[k_idx,pooling_size*r_idx+max_index/pooling_size,pooling_size*c_idx+max_index%pooling_size]=1return out_pooling,max_index_Matirxdef poolwithfunc(in_pooling,W,B,type_func='sigmoid'):k,row,col=numpy.shape(in_pooling)out_pooling=numpy.zeros((k,row,col))for k_idx in range(0,k):for r_idx in range(0,row):for c_idx in range(0,col):out_pooling[k_idx,r_idx,c_idx]=sigmoid(W[k_idx]*in_pooling[k_idx,r_idx,c_idx]+B[k_idx])return out_pooling#out_feature is the out put of convdef backErrorfromPoolToConv(theta,max_index_Matirx,out_feature,pooling_size=2):k1,row,col=numpy.shape(out_feature)error_conv=numpy.zeros((k1,row,col))k2,theta_row,theta_col=numpy.shape(theta)if k1!=k2:raise Exceptionfor idx_k in range(0,k1):for idx_row in range( 0, row):for idx_col in range( 0, col):error_conv[idx_k,idx_row,idx_col]=\max_index_Matirx[idx_k,idx_row,idx_col]*\float(theta[idx_k,idx_row/pooling_size,idx_col/pooling_size])*\difsigmoid(out_feature[idx_k,idx_row,idx_col])return error_convdef backErrorfromConvToInput(theta,inputImage):k1,row,col=numpy.shape(theta)#print "theta",k1,row,coli_row,i_col=numpy.shape(inputImage)if row>i_row or col> i_col:raise Exceptionfilter_row=i_row-row+1filter_col=i_col-col+1detaW=numpy.zeros((k1,filter_row,filter_col))#the same with conv valid in matlabfor k_idx in range(0,k1):for idx_row in range(0,filter_row):for idx_col in range(0,filter_col):subInputMatrix=inputImage[idx_row:idx_row+row,idx_col:idx_col+col]#print "subInputMatrix",numpy.shape(subInputMatrix)#rotate theta 180#print numpy.shape(theta)theta_rotate=numpy.rot90(theta[k_idx,:,:], 2)#print "theta_rotate",theta_rotate(subInputMatrix,theta_rotate)detaW[k_idx,idx_row,idx_col](dotMatrix)detaB=numpy.zeros((k1,1))for k_idx in range(0,k1):detaB[k_idx](theta[k_idx,:,:])return detaW,detaBdef loadMNISTimage(absFilePathandName,datanum=60000):images=open(absFilePathandName,'rb')()index=0magic, numImages , numRows , numColumns = struct.unpack_from('>IIII' , buf , index)print magic, numImages , numRows , numColumnsindex += struct.calcsize('>IIII')if magic != 2051:raise Exceptiondatasize=int(784*datanum)datablock=">"+str(datasize)+"B"#nextmatrix=struct.unpack_from('>47040000B' ,buf, index)nextmatrix=struct.unpack_from(datablock ,buf, index)nextmatrix=numpy.array(nextmatrix)/255.0#nextmatrix=nextmatrix.reshape(numImages,numRows,numColumns)#nextmatrix=nextmatrix.reshape(datanum,1,numRows*numColumns)nextmatrix=nextmatrix.reshape(datanum,1,numRows,numColumns)return nextmatrix, numImagesdef loadMNISTlabels(absFilePathandName,datanum=60000):labels=open(absFilePathandName,'rb')()index=0magic, numLabels = struct.unpack_from('>II' , buf , index)print magic, numLabelsindex += struct.calcsize('>II')if magic != 2049:raise Exceptiondatablock=">"+str(datanum)+"B"#nextmatrix=struct.unpack_from('>60000B' ,buf, index)nextmatrix=struct.unpack_from(datablock ,buf, index)nextmatrix=numpy.array(nextmatrix)return nextmatrix, numLabelsdef simpleCNN(numofFilter,filter_size,pooling_size=2,maxIter=1000,imageNum=500):decayRate=0.01MNISTimage,num1=loadMNISTimage("F:\Machine Learning\UFLDL\data\common\\train-images-idx3-ubyte",imageNum)print num1row,col=numpy.shape(MNISTimage[0,0,:,:])out_Di=numofFilter*((row-filter_size+1)/pooling_size)*((col-filter_size+1)/pooling_size)MLP=BMNN2.MuiltilayerANN(1,[128],out_Di,10,maxIter)MLP.setTrainDataNum(imageNum)MLP.loadtrainlabel("F:\Machine Learning\UFLDL\data\common\\train-labels-idx1-ubyte")MLP.initialweights()#MLP.printWeightMatrix()rng = numpy.random.RandomState(23455)W_shp = (numofFilter, filter_size, filter_size)W_bound = (numofFilter * filter_size * filter_size)W_k=rng.uniform(low=-1.0 / W_bound,high=1.0 / W_bound,size=W_shp)B_shp = (numofFilter,)B= numpy.asarray(rng.uniform(low=-.5, high=.5, size=B_shp))cIter=0while cIter。
dense层的目的,是将前面提取的特征,在dense经过非线性变化,提取这些特征之间的关联,最后映射到输出空间上。
理论上,一层dense足够,但这只是理论上,因为你不清楚这一层dense需要多少个节点的,也不知道需要多少次的训练,加更多的dense,能更快的收敛。
你自行搭建的神经网络模型,权值和阈值仍然是要通过训练得到的。初始化后,将BP算法加到这个模型上,不断调整权值。可以先用神经网络工具箱训练好一个网络,再将权值和阈值导出。
{1,1}=W1; {2,1}=W2; net.b{1}=B1; net.b{2}=B2;注意要反过来,如果是导出的话。
BP神经网络,指的是用了“BP算法”进行训练的“多层感知器模型”。
多层感知器(MLP,MultilayerPerceptron)是一种前馈人工神经网络模型,其将输入的多个数据集映射到单一的输出的数据集上,可以解决任何线性不可分问题。不要把算法和网络搞混了。