9.卷积神经网络操作

9.1 填充与步幅计算

import torch 
from torch import nndef comp_conv2d(conv2d,X):X=X.reshape((1,1)+X.shape)#(BS,CHANNEL,H,W)Y=conv2d(X)return Y.reshape(Y.shape[2:])#这个时候在把前面维度拿掉
#测试样例1
conv2d=nn.Conv2d(1,1,kernel_size=3,padding=1)
X=torch.rand(size=(8,8))
print(comp_conv2d(conv2d,X).shape)#8+2*1-(3-1)=8
#测试样例2
conv2d=nn.Conv2d(1,1,kernel_size=(5,3),padding=(2,1))
X=torch.rand(size=(8,8))
print(comp_conv2d(conv2d,X).shape)#8+2*2-(5-1)=8|8+2*1-(3-1)=8
#测试样例3
conv2d=nn.Conv2d(1,1,kernel_size=(5,3),padding=(2,1),stride=2)
X=torch.rand(size=(8,8))
print(comp_conv2d(conv2d,X).shape)#8+2*2-5/2+1=4|8+2*1-3/2+1=4,向下取值floor
#
conv2d=nn.Conv2d(1,1,kernel_size=(3,5),padding=(0,1),stride=(3,4))
X=torch.rand(size=(8,8))
print(comp_conv2d(conv2d,X).shape)#8+2*0-(3-1)/3+1=2|8+2*1-5+1/4+1=2

9.2 多输入通道计算

#多个输入通道
import torch
from torch import nn
import matplotlib.pyplot as plt
def corr2d(X,k):#计算二维互相关运算h,w=k.shape#卷积核的长和宽Y=torch.zeros((X.shape[0]-h+1,X.shape[1]-w+1))#创建(X-H+1,X-W+1)的全零矩阵for i in range(Y.shape[0]):for j in range(Y.shape[1]):Y[i,j]=(X[i:i+h,j:j+w]*k).sum()#不指定dim的话是全部都加载一起#print(X[i:i+h,j:j+w]*k)return Y
#首先需要把每一个输入通道与每一个卷积核计算corr然后把结果取sum()
def corr2d_multi_in(X,K):return sum(corr2d(x,k) for x,k in zip(X,K))
X = torch.tensor([[[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]],[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]]])#2*3*3
K = torch.tensor([[[0.0, 1.0], [2.0, 3.0]],[[1.0, 2.0], [3.0, 4.0]]])
corr2d_multi_in(X,K)

9.3 多输出通道计算

#多输出通道
import torch
from torch import nn
import matplotlib.pyplot as plt
def corr2d(X,k):#计算二维互相关运算h,w=k.shape#卷积核的长和宽Y=torch.zeros((X.shape[0]-h+1,X.shape[1]-w+1))#创建(X-H+1,X-W+1)的全零矩阵for i in range(Y.shape[0]):for j in range(Y.shape[1]):Y[i,j]=(X[i:i+h,j:j+w]*k).sum()#不指定dim的话是全部都加载一起#print(X[i:i+h,j:j+w]*k)return Y
def corr2d_multi_in(X,K):return sum(corr2d(x,k) for x,k in zip(X,K))
#首先需要把每一个输入通道与每一个卷积核计算corr然后把结果取sum()
def corr2d_multi_in_out(X,K):return torch.stack([corr2d_multi_in(X,k) for k in K],0)#return sum(corr2d(x,k) for x,k in zip(X,K))'
K = torch.tensor([[[0.0, 1.0], [2.0, 3.0]],[[1.0, 2.0], [3.0, 4.0]]])
K = torch.stack((K,K+1,K+2),0)
print(K.shape)
X = torch.tensor([[[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]],[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]]])#2*3*3
#将X的每一个通道与每一个卷积核做计算之后相加得到一个2x2的特征，然后由于有3个卷积，所以stack在一起就变成[3,2,2]的结果了
corr2d_multi_in_out(X,K)

9.4 多输出通道计算(1x1卷积)

#1x1的卷积核测试
import torch
from torch import nn
import matplotlib.pyplot as plt
def corr2d(X,k):#计算二维互相关运算h,w=k.shape#卷积核的长和宽Y=torch.zeros((X.shape[0]-h+1,X.shape[1]-w+1))#创建(X-H+1,X-W+1)的全零矩阵for i in range(Y.shape[0]):for j in range(Y.shape[1]):Y[i,j]=(X[i:i+h,j:j+w]*k).sum()#不指定dim的话是全部都加载一起#print(X[i:i+h,j:j+w]*k)return Y
def corr2d_multi_in(X,K):return sum(corr2d(x,k) for x,k in zip(X,K))
#首先需要把每一个输入通道与每一个卷积核计算corr然后把结果取sum()
def corr2d_multi_in_out(X,K):return torch.stack([corr2d_multi_in(X,k) for k in K],0)#return sum(corr2d(x,k) for x,k in zip(X,K))'
def corr2d_multi_in_out_1x1(X,K):c_i,h,w=X.shape#[2,2,2]c_o=K.shape[0]#[3,2,2,2]X=X.reshape((c_i,h*w))#flatten()K=K.reshape((c_o,c_i))Y=torch.matmul(K,X)print(Y.shape)return Y.reshape((c_o,h,w))
X = torch.normal(0, 1, (3, 3, 3))
K = torch.normal(0, 1, (2, 3, 1, 1))#K.shape[0]是我自己设定的
Y1=corr2d_multi_in_out_1x1(X,K)
Y2=corr2d_multi_in_out(X,K)
assert float(torch.abs(Y1-Y2).sum())<1e-6#检查Y1==Y2
torch.isclose(Y1,Y2,rtol=1e-6,atol=0.0)

9.5 池化操作

import torch
from torch import nn
import matplotlib.pyplot as pltdef pool2d(X,pool_size,mode='max'):ph,pw=pool_sizeY=torch.zeros((X.shape[0]-ph+1,X.shape[1]-pw+1))for i in range(Y.shape[0]):for j in range(Y.shape[1]):if mode=='max':Y[i,j]=X[i:i+ph,j:j+pw].max()elif mode=='avg':Y[i,j]=X[i:i+ph,j:j+pw].mean()return Y
X = torch.tensor([[0.0, 1.0, 2.0], [3.0, 4.0, 5.0], [6.0, 7.0, 8.0]])
print(pool2d(X, (2, 2),mode='max'))
print(pool2d(X, (2, 2),mode='avg'))