Python3 CNN中卷积和池化的实现--限制为二维输入
1 # -*- coding: utf-8 -*- 2 """ 3 Created on Wed Jan 31 14:10:03 2018 4 5 @author: markli 6 """ 7 import numpy as np; 8 9 def ReLU(x): 10 return max(0,x); 11 12 class CovolutionLayer: 13 """ 14 卷积层包含卷积和汇合两步操作 15 """ 16 def __init__(self,filters,padding,pooling,action_fun=ReLU): 17 """ 18 filters 包含 filter_size filter_num filter_stride 19 filter_size 过滤器(卷积核)大小 [行,列] 20 filter_num 过滤器(卷积核)的个数 21 filter_stride 卷积步长 22 padding 填充的大小 填充0值 23 pooling 包含 pooling_size pooling_stride pooling_classic 24 pooling_size 池化时的大小 [行,列] 池化矩阵为方阵 25 pooling_stride 池化时的步长 一般情况下大小等于pooling_size 26 pooling_classic 汇合类型 类型包括 最大值汇合 max, 平均值汇合 average 27 action_fun 卷积操作的激活函数 28 """ 29 self.f_size = (filters[0],filters[1]); 30 self.f_num = filters[2]; 31 self.f_stride = filters[3]; 32 self.padding = padding; 33 self.p_size = (pooling[0],pooling[1]); 34 self.p_stride = pooling[2] 35 self.p_classic = pooling[3]; 36 self.action_fun = action_fun; 37 38 self.weights = []; 39 for i in range(self.f_num): 40 weight = np.random.randn(self.f_size[0],self.f_size[1]); 41 self.weights.append(weight); 42 43 self.biase = np.random.randn(self.f_num); 44 45 def Convolution(self,X): 46 """ 47 X 为二维数组 48 """ 49 #获得初始X的形状 50 n,m = X.shape; 51 #填充 52 if(self.padding != 0): 53 ones = np.zeros((n+2*self.padding,m+2*self.padding)); 54 ones[self.padding:self.padding+n,self.padding:self.padding+m] = X; 55 X = ones; 56 57 #获得填充后X的形状 58 n,m = X.shape; 59 60 #求得卷积操作降维后的层的大小 61 t = int((n-self.f_size[0])/ self.f_stride) + 1; #行数 62 l = int((m-self.f_size[1]) / self.f_stride) + 1; #列数 63 64 #求得池化后降维的大小 65 t_p = int((t-self.p_size[0]) / self.p_stride) + 1; #池化层的行数 66 l_p = int((l-self.p_size[1]) / self.p_stride) + 1; #池化层的列数 67 68 69 70 self.convs = []; 71 self.pools = [] 72 for k in range(self.f_num): 73 conv = np.ones((t,l)); 74 pool = np.ones((t_p,l_p)); 75 row = 0; 76 #卷积 77 for i in range(l): 78 col = 0; 79 for j in range(t): 80 temp = X[row:row+self.f_size[0],col:col+self.f_size[1]]; 81 z = np.sum(np.multiply(self.weights[k],temp)) + self.biase[k]; 82 a = self.action_fun(z); 83 conv[i][j] = a; 84 col = col + self.f_stride; 85 86 row = row + self.f_stride; 87 88 self.convs.append(conv); 89 90 #池化 91 row = 0; 92 for i in range(t_p): 93 col = 0; 94 for j in range(l_p): 95 temp = conv[row:row+self.p_size[0],col:col+self.p_size[1]]; 96 if(self.p_classic == "average"): 97 pool[i][j] = np.sum(temp) / (self.p_size[0] * self.p_size[1]); 98 else: 99 pool[i][j] = np.max(temp); 100 101 col = col + self.p_stride; 102 103 row = row + self.p_stride; 104 105 self.pools.append(pool); 106 107 X = np.array([[18,54,51,239,244,188], 108 [55,121,75,78,95,88], 109 [35,24,104,113,109,221], 110 [3,154,104,235,25,130], 111 [15,253,225,159,78,233], 112 [68,85,180,214,215,0]]); 113 114 #X 归一化处理 115 X = (X - np.sum(X)/36) / np.max(X); 116 #print(X.shape) 117 118 con = CovolutionLayer([3,3,2,1],1,[2,2,2,"max"],); 119 con.Convolution(X); 120 print(con.pools);
目前只能实现二维的操作，三维的实现还没想好如何存储。卷积神经网络的存储和计算是真的很复杂，过段时间想好了在实现。
